agnt5 0.2.8a13__cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

agnt5/agent.py

@@ -0,0 +1,1774 @@
+"""Agent component implementation for AGNT5 SDK.
+
+Provides simple agent with external LLM integration and tool orchestration.
+Future: Platform-backed agents with durable execution and multi-agent coordination.
+"""
+
+from __future__ import annotations
+
+import functools
+import json
+import logging
+import time
+from typing import Any, Callable, Dict, List, Optional, Union
+
+from .context import Context, get_current_context, set_current_context
+from . import lm
+from .lm import GenerateRequest, GenerateResponse, LanguageModel, Message, ModelConfig, ToolDefinition
+from .tool import Tool, ToolRegistry
+from ._telemetry import setup_module_logger
+from .exceptions import WaitingForUserInputException
+
+logger = setup_module_logger(__name__)
+
+# Global agent registry
+_AGENT_REGISTRY: Dict[str, "Agent"] = {}
+
+
+class AgentContext(Context):
+    """
+    Context for agent execution with conversation state management.
+
+    Extends base Context with:
+    - State management via EntityStateManager
+    - Conversation history persistence
+    - Context inheritance (child agents share parent's state)
+
+    Three initialization modes:
+    1. Standalone: Creates own state manager (playground testing)
+    2. Inherit WorkflowContext: Shares parent's state manager
+    3. Inherit parent AgentContext: Shares parent's state manager
+
+    Example:
+        ```python
+        # Standalone agent with conversation history
+        ctx = AgentContext(run_id="session-1", agent_name="tutor")
+        result = await agent.run("Hello", context=ctx)
+        result = await agent.run("Continue", context=ctx)  # Remembers previous message
+
+        # Agent in workflow - shares workflow state
+        @workflow
+        async def research_workflow(ctx: WorkflowContext):
+            agent_result = await research_agent.run("Find AI trends", context=ctx)
+            # Agent has access to workflow state via inherited context
+        ```
+    """
+
+    def __init__(
+        self,
+        run_id: str,
+        agent_name: str,
+        session_id: Optional[str] = None,
+        state_manager: Optional[Any] = None,
+        parent_context: Optional[Context] = None,
+        attempt: int = 0,
+        runtime_context: Optional[Any] = None,
+    ):
+        """
+        Initialize agent context.
+
+        Args:
+            run_id: Unique execution identifier
+            agent_name: Name of the agent
+            session_id: Session identifier for conversation history (default: run_id)
+            state_manager: Optional state manager (for context inheritance)
+            parent_context: Parent context to inherit state from
+            attempt: Retry attempt number
+            runtime_context: RuntimeContext for trace correlation
+        """
+        super().__init__(run_id, attempt, runtime_context)
+
+        self._agent_name = agent_name
+        self._session_id = session_id or run_id
+        self.parent_context = parent_context  # Store for context chain traversal
+
+        # Determine state adapter based on parent context
+        from .entity import EntityStateAdapter, _get_state_adapter
+
+        if state_manager:
+            # Explicit state adapter provided (parameter name kept for backward compat)
+            self._state_adapter = state_manager
+        elif parent_context:
+            # Try to inherit state adapter from parent
+            try:
+                # Check if parent is WorkflowContext or AgentContext
+                if hasattr(parent_context, '_workflow_entity'):
+                    # WorkflowContext - get state adapter from worker context
+                    self._state_adapter = _get_state_adapter()
+                elif hasattr(parent_context, '_state_adapter'):
+                    # Parent AgentContext - share state adapter
+                    self._state_adapter = parent_context._state_adapter
+                elif hasattr(parent_context, '_state_manager'):
+                    # Backward compatibility: parent has old _state_manager
+                    self._state_adapter = parent_context._state_manager
+                else:
+                    # FunctionContext or base Context - create new state adapter
+                    self._state_adapter = EntityStateAdapter()
+            except RuntimeError as e:
+                # _get_state_adapter() failed (not in worker context) - create standalone
+                self._state_adapter = EntityStateAdapter()
+        else:
+            # Try to get from worker context first
+            try:
+                self._state_adapter = _get_state_adapter()
+            except RuntimeError as e:
+                # Standalone - create new state adapter
+                self._state_adapter = EntityStateAdapter()
+
+        # Conversation key for state storage (used for in-memory state)
+        self._conversation_key = f"agent:{agent_name}:{self._session_id}:messages"
+        # Entity key for database persistence (without :messages suffix to match API expectations)
+        self._entity_key = f"agent:{agent_name}:{self._session_id}"
+
+        # Determine storage mode: "workflow" if parent is WorkflowContext, else "standalone"
+        self._storage_mode = "standalone"  # Default mode
+        self._workflow_entity = None
+
+        if parent_context and hasattr(parent_context, '_workflow_entity'):
+            # Agent is running within a workflow - store conversation in workflow state
+            self._storage_mode = "workflow"
+            self._workflow_entity = parent_context._workflow_entity
+            logger.debug(f"Agent '{agent_name}' using workflow storage mode (workflow entity: {self._workflow_entity.key})")
+
+    @property
+    def state(self):
+        """
+        Get state interface for agent state management.
+
+        Note: This is a simplified in-memory state interface for agent-specific data.
+        Conversation history is managed separately via get_conversation_history() and
+        save_conversation_history() which use the Rust-backed persistence layer.
+
+        Returns:
+            Dict-like object for state operations
+
+        Example:
+            # Store agent-specific data (in-memory only)
+            ctx.state["research_results"] = data
+            ctx.state["iteration_count"] = 5
+        """
+        # Simple dict-based state for agent-specific data
+        # This is in-memory only and not persisted to platform
+        if not hasattr(self, '_agent_state'):
+            self._agent_state = {}
+        return self._agent_state
+
+    @property
+    def session_id(self) -> str:
+        """Get session identifier for this agent context."""
+        return self._session_id
+
+    async def get_conversation_history(self) -> List[Message]:
+        """
+        Retrieve conversation history from state, loading from database if needed.
+
+        Uses the EntityStateAdapter which delegates to Rust core for cache-first loading.
+        If running within a workflow, loads from workflow entity state instead.
+
+        Returns:
+            List of Message objects from conversation history
+        """
+        if self._storage_mode == "workflow":
+            return await self._load_from_workflow_state()
+        else:
+            return await self._load_from_entity_storage()
+
+    async def _load_from_workflow_state(self) -> List[Message]:
+        """Load conversation history from workflow entity state."""
+        key = f"agent.{self._agent_name}"
+        agent_data = self._workflow_entity.state.get(key, {})
+        messages_data = agent_data.get("messages", [])
+
+        # Convert dict representations back to Message objects
+        return self._convert_dicts_to_messages(messages_data)
+
+    async def _load_from_entity_storage(self) -> List[Message]:
+        """Load conversation history from AgentSession entity (standalone mode)."""
+        entity_type = "AgentSession"
+        entity_key = self._entity_key
+
+        # Load session data via adapter (Rust handles cache + platform load)
+        session_data = await self._state_adapter.load_state(entity_type, entity_key)
+
+        # Extract messages from session object
+        if isinstance(session_data, dict) and "messages" in session_data:
+            # New format with session metadata
+            messages_data = session_data["messages"]
+        elif isinstance(session_data, list):
+            # Old format - just messages array
+            messages_data = session_data
+        else:
+            # No messages found
+            messages_data = []
+
+        # Convert dict representations back to Message objects
+        return self._convert_dicts_to_messages(messages_data)
+
+    def _convert_dicts_to_messages(self, messages_data: list) -> List[Message]:
+        """Convert list of message dicts to Message objects."""
+        messages = []
+        for msg_dict in messages_data:
+            if isinstance(msg_dict, dict):
+                role = msg_dict.get("role", "user")
+                content = msg_dict.get("content", "")
+                if role == "user":
+                    messages.append(Message.user(content))
+                elif role == "assistant":
+                    messages.append(Message.assistant(content))
+                else:
+                    # Generic message - create with MessageRole enum
+                    from .lm import MessageRole
+                    msg_role = MessageRole(role) if role in ("user", "assistant", "system") else MessageRole.USER
+                    msg = Message(role=msg_role, content=content)
+                    messages.append(msg)
+            else:
+                # Already a Message object
+                messages.append(msg_dict)
+
+        return messages
+
+    async def save_conversation_history(self, messages: List[Message]) -> None:
+        """
+        Save conversation history to state and persist to database.
+
+        Uses the EntityStateAdapter which delegates to Rust core for version-checked saves.
+        If running within a workflow, saves to workflow entity state instead.
+
+        Args:
+            messages: List of Message objects to persist
+        """
+        if self._storage_mode == "workflow":
+            await self._save_to_workflow_state(messages)
+        else:
+            await self._save_to_entity_storage(messages)
+
+    async def _save_to_workflow_state(self, messages: List[Message]) -> None:
+        """Save conversation history to workflow entity state."""
+        # Convert Message objects to dict for JSON serialization
+        messages_data = []
+        for msg in messages:
+            messages_data.append({
+                "role": msg.role.value if hasattr(msg.role, 'value') else str(msg.role),
+                "content": msg.content,
+                "timestamp": time.time()
+            })
+
+        # Build agent data structure
+        key = f"agent.{self._agent_name}"
+        current_data = self._workflow_entity.state.get(key, {})
+        now = time.time()
+
+        agent_data = {
+            "session_id": self._session_id,
+            "agent_name": self._agent_name,
+            "created_at": current_data.get("created_at", now),
+            "last_message_time": now,
+            "message_count": len(messages_data),
+            "messages": messages_data,
+            "metadata": getattr(self, '_custom_metadata', {})
+        }
+
+        # Store in workflow state (WorkflowEntity handles persistence)
+        self._workflow_entity.state.set(key, agent_data)
+        logger.info(f"Saved conversation to workflow state: {key} ({len(messages_data)} messages)")
+
+    async def _save_to_entity_storage(self, messages: List[Message]) -> None:
+        """Save conversation history to AgentSession entity (standalone mode)."""
+        # Convert Message objects to dict for JSON serialization
+        messages_data = []
+        for msg in messages:
+            messages_data.append({
+                "role": msg.role.value if hasattr(msg.role, 'value') else str(msg.role),
+                "content": msg.content,
+                "timestamp": time.time()  # Add timestamp for each message
+            })
+
+        entity_type = "AgentSession"
+        entity_key = self._entity_key
+
+        # Load current state with version for optimistic locking
+        current_state, current_version = await self._state_adapter.load_with_version(
+            entity_type, entity_key
+        )
+
+        # Build session object with metadata
+        now = time.time()
+
+        # Get custom metadata from instance variable or preserve from loaded state
+        custom_metadata = getattr(self, '_custom_metadata', current_state.get("metadata", {}))
+
+        session_data = {
+            "session_id": self._session_id,
+            "agent_name": self._agent_name,
+            "created_at": current_state.get("created_at", now),  # Preserve existing or set new
+            "last_message_time": now,
+            "message_count": len(messages_data),
+            "messages": messages_data,
+            "metadata": custom_metadata  # Save custom metadata
+        }
+
+        # Save to platform via adapter (Rust handles optimistic locking)
+        try:
+            new_version = await self._state_adapter.save_state(
+                entity_type,
+                entity_key,
+                session_data,
+                current_version
+            )
+            logger.info(
+                f"Persisted conversation history: {entity_key} (version {current_version} -> {new_version})"
+            )
+        except Exception as e:
+            logger.error(f"Failed to persist conversation history to database: {e}")
+            # Don't fail - conversation is still in memory for this execution
+
+    async def get_metadata(self) -> Dict[str, Any]:
+        """
+        Get conversation session metadata.
+
+        Returns session metadata including:
+        - created_at: Timestamp of first message (float, Unix timestamp)
+        - last_activity: Timestamp of last message (float, Unix timestamp)
+        - message_count: Number of messages in conversation (int)
+        - custom: Dict of user-provided custom metadata
+
+        Returns:
+            Dictionary with metadata. If no conversation exists yet, returns defaults.
+
+        Example:
+            ```python
+            metadata = await context.get_metadata()
+            print(f"Session created: {metadata['created_at']}")
+            print(f"User ID: {metadata['custom'].get('user_id')}")
+            ```
+        """
+        if self._storage_mode == "workflow":
+            return await self._get_metadata_from_workflow()
+        else:
+            return await self._get_metadata_from_entity()
+
+    async def _get_metadata_from_workflow(self) -> Dict[str, Any]:
+        """Get metadata from workflow entity state."""
+        key = f"agent.{self._agent_name}"
+        agent_data = self._workflow_entity.state.get(key, {})
+
+        if not agent_data:
+            # No conversation exists yet - return defaults
+            return {
+                "created_at": None,
+                "last_activity": None,
+                "message_count": 0,
+                "custom": getattr(self, '_custom_metadata', {})
+            }
+
+        messages = agent_data.get("messages", [])
+        return {
+            "created_at": agent_data.get("created_at"),
+            "last_activity": agent_data.get("last_message_time"),
+            "message_count": len(messages),
+            "custom": agent_data.get("metadata", {})
+        }
+
+    async def _get_metadata_from_entity(self) -> Dict[str, Any]:
+        """Get metadata from AgentSession entity (standalone mode)."""
+        entity_type = "AgentSession"
+        entity_key = self._entity_key
+
+        # Load session data
+        session_data = await self._state_adapter.load_state(entity_type, entity_key)
+
+        if not session_data:
+            # No conversation exists yet - return defaults
+            return {
+                "created_at": None,
+                "last_activity": None,
+                "message_count": 0,
+                "custom": getattr(self, '_custom_metadata', {})
+            }
+
+        messages = session_data.get("messages", [])
+
+        # Derive timestamps from messages if available
+        created_at = session_data.get("created_at")
+        last_activity = session_data.get("last_message_time")
+
+        return {
+            "created_at": created_at,
+            "last_activity": last_activity,
+            "message_count": len(messages),
+            "custom": session_data.get("metadata", {})
+        }
+
+    def update_metadata(self, **kwargs) -> None:
+        """
+        Update custom session metadata.
+
+        Metadata will be persisted alongside conversation history on next save.
+        Use this to store application-specific data like user_id, preferences, etc.
+
+        Args:
+            **kwargs: Key-value pairs to store as metadata
+
+        Example:
+            ```python
+            # Store user identification and preferences
+            context.update_metadata(
+                user_id="user-123",
+                subscription_tier="premium",
+                preferences={"theme": "dark", "language": "en"}
+            )
+
+            # Later retrieve it
+            metadata = await context.get_metadata()
+            user_id = metadata["custom"]["user_id"]
+            ```
+
+        Note:
+            - Metadata is merged with existing metadata (doesn't replace)
+            - Changes persist on next save_conversation_history() call
+            - Use simple JSON-serializable types (str, int, float, dict, list)
+        """
+        if not hasattr(self, '_custom_metadata'):
+            self._custom_metadata = {}
+        self._custom_metadata.update(kwargs)
+
+
+class Handoff:
+    """Configuration for agent-to-agent handoff.
+
+    Handoffs enable one agent to delegate control to another specialized agent,
+    following the pattern popularized by LangGraph and OpenAI Agents SDK.
+
+    The handoff is exposed to the LLM as a tool named 'transfer_to_{agent_name}'
+    that allows explicit delegation with conversation history.
+
+    Example:
+        ```python
+        specialist = Agent(name="specialist", ...)
+
+        # Simple: Pass agent directly (auto-wrapped with defaults)
+        coordinator = Agent(
+            name="coordinator",
+            handoffs=[specialist]  # Agent auto-converted to Handoff
+        )
+
+        # Advanced: Use Handoff for custom configuration
+        coordinator = Agent(
+            name="coordinator",
+            handoffs=[
+                Handoff(
+                    agent=specialist,
+                    description="Custom description for LLM",
+                    tool_name="custom_transfer_name",
+                    pass_full_history=False
+                )
+            ]
+        )
+        ```
+    """
+
+    def __init__(
+        self,
+        agent: "Agent",
+        description: Optional[str] = None,
+        tool_name: Optional[str] = None,
+        pass_full_history: bool = True,
+    ):
+        """Initialize handoff configuration.
+
+        Args:
+            agent: Target agent to hand off to
+            description: Description shown to LLM (defaults to agent instructions)
+            tool_name: Custom tool name (defaults to 'transfer_to_{agent_name}')
+            pass_full_history: Whether to pass full conversation history to target agent
+        """
+        self.agent = agent
+        self.description = description or agent.instructions or f"Transfer to {agent.name}"
+        self.tool_name = tool_name or f"transfer_to_{agent.name}"
+        self.pass_full_history = pass_full_history
+
+
+def handoff(
+    agent: "Agent",
+    description: Optional[str] = None,
+    tool_name: Optional[str] = None,
+    pass_full_history: bool = True,
+) -> Handoff:
+    """Create a handoff configuration for agent-to-agent delegation.
+
+    This is a convenience function for creating Handoff instances with a clean API.
+
+    Args:
+        agent: Target agent to hand off to
+        description: Description shown to LLM
+        tool_name: Custom tool name
+        pass_full_history: Whether to pass full conversation history
+
+    Returns:
+        Handoff configuration
+
+    Example:
+        ```python
+        from agnt5 import Agent, handoff
+
+        research_agent = Agent(name="researcher", ...)
+        writer_agent = Agent(name="writer", ...)
+
+        coordinator = Agent(
+            name="coordinator",
+            handoffs=[
+                handoff(research_agent, "Transfer for research tasks"),
+                handoff(writer_agent, "Transfer for writing tasks"),
+            ]
+        )
+        ```
+    """
+    return Handoff(
+        agent=agent,
+        description=description,
+        tool_name=tool_name,
+        pass_full_history=pass_full_history,
+    )
+
+
+class AgentRegistry:
+    """Registry for agents."""
+
+    @staticmethod
+    def register(agent: "Agent") -> None:
+        """Register an agent."""
+        if agent.name in _AGENT_REGISTRY:
+            logger.warning(f"Overwriting existing agent '{agent.name}'")
+        _AGENT_REGISTRY[agent.name] = agent
+
+    @staticmethod
+    def get(name: str) -> Optional["Agent"]:
+        """Get agent by name."""
+        return _AGENT_REGISTRY.get(name)
+
+    @staticmethod
+    def all() -> Dict[str, "Agent"]:
+        """Get all registered agents."""
+        return _AGENT_REGISTRY.copy()
+
+    @staticmethod
+    def clear() -> None:
+        """Clear all registered agents."""
+        _AGENT_REGISTRY.clear()
+
+
+class AgentResult:
+    """Result from agent execution."""
+
+    def __init__(
+        self,
+        output: str,
+        tool_calls: List[Dict[str, Any]],
+        context: Context,
+        handoff_to: Optional[str] = None,
+        handoff_metadata: Optional[Dict[str, Any]] = None,
+    ):
+        self.output = output
+        self.tool_calls = tool_calls
+        self.context = context
+        self.handoff_to = handoff_to  # Name of agent that was handed off to
+        self.handoff_metadata = handoff_metadata or {}  # Additional handoff info
+
+
+class Agent:
+    """Autonomous LLM-driven agent with tool orchestration.
+
+    Current features:
+    - LLM integration (OpenAI, Anthropic, etc.)
+    - Tool selection and execution
+    - Multi-turn reasoning
+    - Context and state management
+
+    Future enhancements:
+    - Durable execution with checkpointing
+    - Multi-agent coordination
+    - Platform-backed tool execution
+    - Streaming responses
+
+    Example:
+        ```python
+        from agnt5 import Agent, tool, Context
+
+        @tool(auto_schema=True)
+        async def search_web(ctx: Context, query: str) -> List[Dict]:
+            # Search implementation
+            return [{"title": "Result", "url": "..."}]
+
+        # Simple usage with model string
+        agent = Agent(
+            name="researcher",
+            model="openai/gpt-4o-mini",
+            instructions="You are a research assistant.",
+            tools=[search_web],
+            temperature=0.7
+        )
+
+        result = await agent.run("What are the latest AI trends?")
+        print(result.output)
+        ```
+    """
+
+    def __init__(
+        self,
+        name: str,
+        model: Any,  # Can be string like "openai/gpt-4o-mini" OR LanguageModel instance
+        instructions: str,
+        tools: Optional[List[Any]] = None,
+        handoffs: Optional[List[Union["Agent", Handoff]]] = None,  # Accept Agent or Handoff instances
+        temperature: float = 0.7,
+        max_tokens: Optional[int] = None,
+        top_p: Optional[float] = None,
+        model_config: Optional[ModelConfig] = None,
+        max_iterations: int = 10,
+        model_name: Optional[str] = None,  # For backwards compatibility with tests
+    ):
+        """Initialize agent.
+
+        Args:
+            name: Agent name/identifier
+            model: Model string with provider prefix (e.g., "openai/gpt-4o-mini") OR LanguageModel instance
+            instructions: System instructions for the agent
+            tools: List of tools available to the agent (functions, Tool instances, or Agent instances)
+            handoffs: List of handoff configurations - can be Agent instances (auto-wrapped) or Handoff instances for custom config
+            temperature: LLM temperature (0.0 to 1.0)
+            max_tokens: Maximum tokens to generate
+            top_p: Nucleus sampling parameter
+            model_config: Optional advanced configuration (custom endpoints, headers, etc.)
+            max_iterations: Maximum reasoning iterations
+            model_name: Optional model name (for backwards compatibility, used when model is a LanguageModel instance)
+        """
+        self.name = name
+        self.instructions = instructions
+        self.temperature = temperature
+        self.max_tokens = max_tokens
+        self.top_p = top_p
+        self.model_config = model_config
+        self.max_iterations = max_iterations
+
+        # Support both string model names and LanguageModel instances
+        if isinstance(model, str):
+            # New API: model is a string like "openai/gpt-4o-mini"
+            self.model = model
+            self.model_name = model_name or model
+            self._language_model = None  # Will create on demand
+        elif isinstance(model, LanguageModel):
+            # Old API (for tests): model is a LanguageModel instance
+            self._language_model = model
+            self.model = model  # Keep for backwards compatibility
+            self.model_name = model_name or "mock-model"
+        else:
+            raise TypeError(f"model must be a string or LanguageModel instance, got {type(model)}")
+
+        # Normalize handoffs: convert Agent instances to Handoff instances
+        self.handoffs: List[Handoff] = []
+        if handoffs:
+            for handoff_item in handoffs:
+                if isinstance(handoff_item, Agent):
+                    # Auto-wrap Agent in Handoff with sensible defaults
+                    self.handoffs.append(Handoff(agent=handoff_item))
+                    logger.info(f"Auto-wrapped agent '{handoff_item.name}' in Handoff for '{self.name}'")
+                elif isinstance(handoff_item, Handoff):
+                    self.handoffs.append(handoff_item)
+                else:
+                    raise TypeError(f"handoffs must contain Agent or Handoff instances, got {type(handoff_item)}")
+
+        # Build tool registry (includes regular tools, agent-as-tools, and handoff tools)
+        self.tools: Dict[str, Tool] = {}
+        if tools:
+            for tool_item in tools:
+                # Check if it's an Agent instance (agents-as-tools pattern)
+                if isinstance(tool_item, Agent):
+                    agent_tool = tool_item.to_tool()
+                    self.tools[agent_tool.name] = agent_tool
+                    logger.info(f"Added agent '{tool_item.name}' as tool to '{self.name}'")
+                # Check if it's a Tool instance
+                elif isinstance(tool_item, Tool):
+                    self.tools[tool_item.name] = tool_item
+                # Check if it's a decorated function with config
+                elif hasattr(tool_item, "_agnt5_config"):
+                    # Try to get from ToolRegistry first
+                    tool_config = tool_item._agnt5_config
+                    tool_instance = ToolRegistry.get(tool_config.name)
+                    if tool_instance:
+                        self.tools[tool_instance.name] = tool_instance
+                # Otherwise try to look up by function name
+                elif callable(tool_item):
+                    # Try to find in registry by function name
+                    tool_name = tool_item.__name__
+                    tool_instance = ToolRegistry.get(tool_name)
+                    if tool_instance:
+                        self.tools[tool_instance.name] = tool_instance
+
+        # Build handoff tools
+        for handoff_config in self.handoffs:
+            handoff_tool = self._create_handoff_tool(handoff_config)
+            self.tools[handoff_tool.name] = handoff_tool
+            logger.info(f"Added handoff tool '{handoff_tool.name}' to '{self.name}'")
+
+        self.logger = logging.getLogger(f"agnt5.agent.{name}")
+
+        # Cost tracking for this agent's execution
+        self._cumulative_cost_usd = 0.0
+
+        # Define schemas based on the run method signature
+        # Input: user_message (string)
+        self.input_schema = {
+            "type": "object",
+            "properties": {
+                "user_message": {"type": "string"}
+            },
+            "required": ["user_message"]
+        }
+        # Output: AgentResult with output and tool_calls
+        self.output_schema = {
+            "type": "object",
+            "properties": {
+                "output": {"type": "string"},
+                "tool_calls": {
+                    "type": "array",
+                    "items": {"type": "object"}
+                }
+            }
+        }
+
+        # Auto-register agent for discovery
+        AgentRegistry.register(self)
+
+        # Store metadata
+        self.metadata = {
+            "description": instructions,
+            "model": model
+        }
+
+    @property
+    def cumulative_cost_usd(self) -> float:
+        """
+        Get cumulative LLM cost in USD for this agent's execution.
+
+        Returns the total cost of all LLM calls made by this agent instance
+        since it was created.
+
+        Returns:
+            float: Cumulative cost in USD
+
+        Example:
+            ```python
+            result = await agent.run("Analyze this data")
+            print(f"Total cost: ${agent.cumulative_cost_usd:.4f}")
+            ```
+        """
+        return self._cumulative_cost_usd
+
+    def _track_llm_cost(self, response: GenerateResponse, workflow_ctx: Optional[Any] = None) -> None:
+        """
+        Track LLM cost from response and emit cost accrual event.
+
+        Reads cost from OpenTelemetry span attributes (calculated by Rust core)
+        and emits a run.cost.accrued checkpoint event if in workflow context.
+
+        Args:
+            response: LLM response with usage data
+            workflow_ctx: Optional workflow context for checkpoint emission
+        """
+        # Try to get cost from current OpenTelemetry span
+        cost_usd = None
+        try:
+            from opentelemetry import trace
+            span = trace.get_current_span()
+            if span.is_recording():
+                # Cost is set by Rust telemetry in gen_ai.usage.cost attribute
+                attributes = span.attributes or {}
+                cost_usd = attributes.get("gen_ai.usage.cost")
+        except Exception as e:
+            self.logger.debug(f"Could not read cost from span: {e}")
+
+        # Fallback: calculate cost from usage if span doesn't have it
+        # (This shouldn't happen in production, but useful for testing)
+        if cost_usd is None and response.usage:
+            self.logger.debug("Span cost not available, skipping cost tracking")
+            return
+
+        if cost_usd is not None:
+            # Update cumulative cost
+            self._cumulative_cost_usd += cost_usd
+
+            # Emit cost accrual event if in workflow context
+            if workflow_ctx is not None:
+                event_data = {
+                    "cost_usd": cost_usd,
+                    "cumulative_cost_usd": self._cumulative_cost_usd,
+                    "agent_name": self.name,
+                    "model": self.model_name,
+                }
+
+                # Add token usage if available
+                if response.usage:
+                    event_data["input_tokens"] = response.usage.prompt_tokens
+                    event_data["output_tokens"] = response.usage.completion_tokens
+                    event_data["total_tokens"] = response.usage.total_tokens
+
+                workflow_ctx._send_checkpoint("run.cost.accrued", event_data)
+
+            self.logger.debug(
+                f"Agent '{self.name}' cost: ${cost_usd:.6f} "
+                f"(cumulative: ${self._cumulative_cost_usd:.6f})"
+            )
+
+    def to_tool(self, description: Optional[str] = None) -> Tool:
+        """Convert this agent to a Tool that can be used by other agents.
+
+        This enables agents-as-tools pattern where one agent can invoke another
+        agent as if it were a regular tool.
+
+        Args:
+            description: Optional custom description (defaults to agent instructions)
+
+        Returns:
+            Tool instance that wraps this agent
+
+        Example:
+            ```python
+            research_agent = Agent(
+                name="researcher",
+                model="openai/gpt-4o-mini",
+                instructions="You are a research specialist."
+            )
+
+            # Use research agent as a tool for another agent
+            coordinator = Agent(
+                name="coordinator",
+                model="openai/gpt-4o-mini",
+                instructions="Coordinate tasks using specialist agents.",
+                tools=[research_agent.to_tool()]
+            )
+            ```
+        """
+        agent_name = self.name
+
+        # Handler that runs the agent
+        async def agent_tool_handler(ctx: Context, user_message: str) -> str:
+            """Execute agent and return output."""
+            ctx.logger.info(f"Invoking agent '{agent_name}' as tool")
+
+            # Run the agent with the user message
+            result = await self.run(user_message, context=ctx)
+
+            return result.output
+
+        # Create tool with agent's schema
+        tool_description = description or self.instructions or f"Agent: {self.name}"
+
+        agent_tool = Tool(
+            name=self.name,
+            description=tool_description,
+            handler=agent_tool_handler,
+            input_schema=self.input_schema,
+            auto_schema=False,
+        )
+
+        return agent_tool
+
+    def _create_handoff_tool(self, handoff_config: Handoff, current_messages_callback: Optional[Callable] = None) -> Tool:
+        """Create a tool for handoff to another agent.
+
+        Args:
+            handoff_config: Handoff configuration
+            current_messages_callback: Optional callback to get current conversation messages
+
+        Returns:
+            Tool instance that executes the handoff
+        """
+        target_agent = handoff_config.agent
+        tool_name = handoff_config.tool_name
+
+        # Handler that executes the handoff
+        async def handoff_handler(ctx: Context, message: str) -> Dict[str, Any]:
+            """Transfer control to target agent."""
+            ctx.logger.info(
+                f"Handoff from '{self.name}' to '{target_agent.name}': {message}"
+            )
+
+            # If we should pass conversation history, add it to context
+            if handoff_config.pass_full_history:
+                # Get current conversation from the agent's run loop
+                # (This will be set when we detect the handoff in run())
+                conversation_history = getattr(ctx, '_agent_data', {}).get("_current_conversation", [])
+
+                if conversation_history:
+                    ctx.logger.info(
+                        f"Passing {len(conversation_history)} messages to target agent"
+                    )
+                    # Store in context for target agent to optionally use
+                    if not hasattr(ctx, '_agent_data'):
+                        ctx._agent_data = {}
+                    ctx._agent_data["_handoff_conversation_history"] = conversation_history
+
+            # Execute target agent with the message and shared context
+            result = await target_agent.run(message, context=ctx)
+
+            # Store handoff metadata - this signals that a handoff occurred
+            handoff_data = {
+                "_handoff": True,
+                "from_agent": self.name,
+                "to_agent": target_agent.name,
+                "message": message,
+                "output": result.output,
+                "tool_calls": result.tool_calls,
+            }
+
+            if not hasattr(ctx, '_agent_data'):
+                ctx._agent_data = {}
+            ctx._agent_data["_handoff_result"] = handoff_data
+
+            # Return the handoff data (will be detected in run() loop)
+            return handoff_data
+
+        # Create tool with handoff schema
+        handoff_tool = Tool(
+            name=tool_name,
+            description=handoff_config.description,
+            handler=handoff_handler,
+            input_schema={
+                "type": "object",
+                "properties": {
+                    "message": {
+                        "type": "string",
+                        "description": "Message or task to pass to the target agent"
+                    }
+                },
+                "required": ["message"]
+            },
+            auto_schema=False,
+        )
+
+        return handoff_tool
+
+    def _detect_memory_scope(self, context: Optional[Context]) -> tuple[str, str]:
+        """
+        Auto-detect memory scope from context for agent conversation persistence.
+
+        Implements priority logic:
+        1. user_id → user-scoped memory (long-term)
+        2. session_id → session-scoped memory (multi-turn)
+        3. run_id → run-scoped memory (ephemeral)
+
+        Args:
+            context: WorkflowContext or other context with memory scoping fields
+
+        Returns:
+            Tuple of (entity_key, scope) where:
+            - entity_key: e.g., "user:user-456", "session:abc-123", "run:xyz-789"
+            - scope: "user", "session", or "run"
+
+        Example:
+            entity_key, scope = agent._detect_memory_scope(ctx)
+            # If ctx.user_id="user-123": ("user:user-123", "user")
+            # If ctx.session_id="sess-456": ("session:sess-456", "session")
+            # Otherwise: ("run:run-789", "run")
+        """
+        # Extract identifiers from context
+        user_id = getattr(context, 'user_id', None) if context else None
+        session_id = getattr(context, 'session_id', None) if context else None
+        run_id = getattr(context, 'run_id', None) if context else None
+
+        # Priority: user_id > session_id > run_id
+        if user_id:
+            return (f"user:{user_id}", "user")
+        elif session_id and session_id != run_id:  # Explicit session (not defaulting to run_id)
+            return (f"session:{session_id}", "session")
+        elif run_id:
+            return (f"run:{run_id}", "run")
+        else:
+            # Fallback: create ephemeral key
+            import uuid
+            fallback_run_id = f"agent-{self.name}-{uuid.uuid4().hex[:8]}"
+            return (f"run:{fallback_run_id}", "run")
+
+    async def run(
+        self,
+        user_message: str,
+        context: Optional[Context] = None,
+    ) -> AgentResult:
+        """Run agent to completion.
+
+        Args:
+            user_message: User's input message
+            context: Optional context (auto-created if not provided, or read from contextvar)
+
+        Returns:
+            AgentResult with output and execution details
+
+        Example:
+            ```python
+            result = await agent.run("Analyze recent tech news")
+            print(result.output)
+            ```
+        """
+        # Create or adapt context
+        if context is None:
+            # Try to get context from task-local storage (set by workflow/function decorator)
+            context = get_current_context()
+
+        # IMPORTANT: Capture workflow context NOW before we replace it with AgentContext
+        # This allows LM calls inside the agent to emit workflow checkpoints
+        from .workflow import WorkflowContext
+        workflow_ctx = context if isinstance(context, WorkflowContext) else None
+
+        if context is None:
+            # Standalone execution - create AgentContext
+            import uuid
+            run_id = f"agent-{self.name}-{uuid.uuid4().hex[:8]}"
+            context = AgentContext(
+                run_id=run_id,
+                agent_name=self.name,
+            )
+        elif isinstance(context, AgentContext):
+            # Already AgentContext - use as-is
+            pass
+        elif hasattr(context, '_workflow_entity'):
+            # WorkflowContext - create AgentContext that inherits state
+            # Auto-detect memory scope based on user_id/session_id/run_id priority
+            entity_key, scope = self._detect_memory_scope(context)
+
+            import uuid
+            run_id = f"{context.run_id}:agent:{self.name}"
+            # Extract the ID from entity_key (e.g., "session:abc-123" → "abc-123")
+            detected_session_id = entity_key.split(":", 1)[1] if ":" in entity_key else context.run_id
+
+            context = AgentContext(
+                run_id=run_id,
+                agent_name=self.name,
+                session_id=detected_session_id,  # Use auto-detected scope
+                parent_context=context,
+                runtime_context=getattr(context, '_runtime_context', None),  # Inherit trace context
+            )
+        else:
+            # FunctionContext or other - create new AgentContext
+            import uuid
+            run_id = f"{context.run_id}:agent:{self.name}"
+            context = AgentContext(
+                run_id=run_id,
+                agent_name=self.name,
+                runtime_context=getattr(context, '_runtime_context', None),  # Inherit trace context
+            )
+
+        # Emit checkpoint if called within a workflow context
+        if workflow_ctx is not None:
+            workflow_ctx._send_checkpoint("agent.started", {
+                "agent.name": self.name,
+                "agent.model": self.model_name,
+                "agent.tools": list(self.tools.keys()),
+                "agent.max_iterations": self.max_iterations,
+                "user_message": user_message,
+            })
+
+        # NEW: Check if this is a resume from HITL
+        if workflow_ctx and hasattr(workflow_ctx, "_agent_resume_info"):
+            resume_info = workflow_ctx._agent_resume_info
+            if resume_info["agent_name"] == self.name:
+                self.logger.info("Detected HITL resume, calling resume_from_hitl()")
+
+                # Clear resume info to avoid re-entry
+                delattr(workflow_ctx, "_agent_resume_info")
+
+                # Resume from checkpoint (context setup happens inside resume_from_hitl)
+                return await self.resume_from_hitl(
+                    context=workflow_ctx,
+                    agent_context=resume_info["agent_context"],
+                    user_response=resume_info["user_response"],
+                )
+
+        # Set context in task-local storage for automatic propagation to tools and LM calls
+        token = set_current_context(context)
+        try:
+            try:
+                # Load conversation history from state (if AgentContext)
+                if isinstance(context, AgentContext):
+                    messages: List[Message] = await context.get_conversation_history()
+                    # Add new user message
+                    messages.append(Message.user(user_message))
+                    # Save updated conversation
+                    await context.save_conversation_history(messages)
+                else:
+                    # Fallback for non-AgentContext (shouldn't happen with code above)
+                    messages = [Message.user(user_message)]
+
+                # Create span for agent execution with trace linking
+                from ._core import create_span
+
+                with create_span(
+                    self.name,
+                    "agent",
+                    context._runtime_context if hasattr(context, "_runtime_context") else None,
+                    {
+                        "agent.name": self.name,
+                        "agent.model": self.model_name,  # Use model_name (always a string)
+                        "agent.max_iterations": str(self.max_iterations),
+                    },
+                ) as span:
+                    all_tool_calls: List[Dict[str, Any]] = []
+
+                    # Reasoning loop
+                    for iteration in range(self.max_iterations):
+                        # Build tool definitions for LLM
+                        tool_defs = [
+                            ToolDefinition(
+                                name=tool.name,
+                                description=tool.description,
+                                parameters=tool.input_schema,
+                            )
+                            for tool in self.tools.values()
+                        ]
+
+                        # Convert messages to dict format for lm.generate()
+                        messages_dict = []
+                        for msg in messages:
+                            messages_dict.append({
+                                "role": msg.role.value,
+                                "content": msg.content
+                            })
+
+                        # Call LLM
+                        # Check if we have a legacy LanguageModel instance or need to create one
+                        if self._language_model is not None:
+                            # Legacy API: use provided LanguageModel instance
+                            request = GenerateRequest(
+                                model="mock-model",  # Not used by MockLanguageModel
+                                system_prompt=self.instructions,
+                                messages=messages,
+                                tools=tool_defs if tool_defs else [],
+                            )
+                            request.config.temperature = self.temperature
+                            if self.max_tokens:
+                                request.config.max_tokens = self.max_tokens
+                            if self.top_p:
+                                request.config.top_p = self.top_p
+                            response = await self._language_model.generate(request)
+
+                            # Track cost for this LLM call
+                            self._track_llm_cost(response, workflow_ctx)
+                        else:
+                            # New API: model is a string, create internal LM instance
+                            request = GenerateRequest(
+                                model=self.model,
+                                system_prompt=self.instructions,
+                                messages=messages,
+                                tools=tool_defs if tool_defs else [],
+                            )
+                            request.config.temperature = self.temperature
+                            if self.max_tokens:
+                                request.config.max_tokens = self.max_tokens
+                            if self.top_p:
+                                request.config.top_p = self.top_p
+
+                            # Create internal LM instance for generation
+                            # TODO: Use model_config when provided
+                            from .lm import _LanguageModel
+                            provider, model_name = self.model.split('/', 1)
+                            internal_lm = _LanguageModel(provider=provider.lower(), default_model=None)
+                            response = await internal_lm.generate(request)
+
+                            # Track cost for this LLM call
+                            self._track_llm_cost(response, workflow_ctx)
+
+                        # Add assistant response to messages
+                        messages.append(Message.assistant(response.text))
+
+                        # Check if LLM wants to use tools
+                        if response.tool_calls:
+                            self.logger.debug(f"Agent calling {len(response.tool_calls)} tool(s)")
+
+                            # Store current conversation in context for potential handoffs
+                            # Use a simple dict attribute since we don't need full state persistence for this
+                            if not hasattr(context, '_agent_data'):
+                                context._agent_data = {}
+                            context._agent_data["_current_conversation"] = messages
+
+                            # Execute tool calls
+                            tool_results = []
+                            for tool_call in response.tool_calls:
+                                tool_name = tool_call["name"]
+                                tool_args_str = tool_call["arguments"]
+
+                                # Track tool call
+                                all_tool_calls.append(
+                                    {
+                                        "name": tool_name,
+                                        "arguments": tool_args_str,
+                                        "iteration": iteration + 1,
+                                    }
+                                )
+
+                                # Execute tool
+                                try:
+                                    # Parse arguments
+                                    tool_args = json.loads(tool_args_str)
+
+                                    # Get tool
+                                    tool = self.tools.get(tool_name)
+                                    if not tool:
+                                        result_text = f"Error: Tool '{tool_name}' not found"
+                                    else:
+                                        # Execute tool
+                                        result = await tool.invoke(context, **tool_args)
+
+                                        # Check if this was a handoff
+                                        if isinstance(result, dict) and result.get("_handoff"):
+                                            self.logger.info(
+                                                f"Handoff detected to '{result['to_agent']}', "
+                                                f"terminating current agent"
+                                            )
+                                            # Save conversation before returning
+                                            if isinstance(context, AgentContext):
+                                                await context.save_conversation_history(messages)
+                                            # Return immediately with handoff result
+                                            return AgentResult(
+                                                output=result["output"],
+                                                tool_calls=all_tool_calls + result.get("tool_calls", []),
+                                                context=context,
+                                                handoff_to=result["to_agent"],
+                                                handoff_metadata=result,
+                                            )
+
+                                        result_text = json.dumps(result) if result else "null"
+
+                                    tool_results.append(
+                                        {"tool": tool_name, "result": result_text, "error": None}
+                                    )
+
+                                except WaitingForUserInputException as e:
+                                    # HITL PAUSE: Capture agent state and propagate exception
+                                    self.logger.info(f"Agent pausing for user input at iteration {iteration}")
+
+                                    # Serialize messages to dict format
+                                    messages_dict = [
+                                        {"role": msg.role.value, "content": msg.content}
+                                        for msg in messages
+                                    ]
+
+                                    # Enhance exception with agent execution context
+                                    raise WaitingForUserInputException(
+                                        question=e.question,
+                                        input_type=e.input_type,
+                                        options=e.options,
+                                        checkpoint_state=e.checkpoint_state,
+                                        agent_context={
+                                            "agent_name": self.name,
+                                            "iteration": iteration,
+                                            "messages": messages_dict,
+                                            "tool_results": tool_results,
+                                            "pending_tool_call": {
+                                                "name": tool_call["name"],
+                                                "arguments": tool_call["arguments"],
+                                                "tool_call_index": response.tool_calls.index(tool_call),
+                                            },
+                                            "all_tool_calls": all_tool_calls,
+                                            "model_config": {
+                                                "model": self.model,
+                                                "temperature": self.temperature,
+                                                "max_tokens": self.max_tokens,
+                                                "top_p": self.top_p,
+                                            },
+                                        },
+                                    ) from e
+
+                                except Exception as e:
+                                    # Regular tool errors - log and continue
+                                    self.logger.error(f"Tool execution error: {e}")
+                                    tool_results.append(
+                                        {"tool": tool_name, "result": None, "error": str(e)}
+                                    )
+
+                            # Add tool results to conversation
+                            results_text = "\n".join(
+                                [
+                                    f"Tool: {tr['tool']}\nResult: {tr['result']}"
+                                    if tr["error"] is None
+                                    else f"Tool: {tr['tool']}\nError: {tr['error']}"
+                                    for tr in tool_results
+                                ]
+                            )
+                            messages.append(Message.user(f"Tool results:\n{results_text}\n\nPlease provide your final answer based on these results."))
+
+                            # Continue loop for agent to process results
+
+                        else:
+                            # No tool calls - agent is done
+                            self.logger.debug(f"Agent completed after {iteration + 1} iterations")
+                            # Save conversation before returning
+                            if isinstance(context, AgentContext):
+                                await context.save_conversation_history(messages)
+
+                            # Emit completion checkpoint
+                            if workflow_ctx:
+                                workflow_ctx._send_checkpoint("agent.completed", {
+                                    "agent.name": self.name,
+                                    "agent.iterations": iteration + 1,
+                                    "agent.tool_calls_count": len(all_tool_calls),
+                                    "output_length": len(response.text),
+                                })
+
+                            return AgentResult(
+                                output=response.text,
+                                tool_calls=all_tool_calls,
+                                context=context,
+                            )
+
+                    # Max iterations reached
+                    self.logger.warning(f"Agent reached max iterations ({self.max_iterations})")
+                    final_output = messages[-1].content if messages else "No output generated"
+                    # Save conversation before returning
+                    if isinstance(context, AgentContext):
+                        await context.save_conversation_history(messages)
+
+                    # Emit completion checkpoint with max iterations flag
+                    if workflow_ctx:
+                        workflow_ctx._send_checkpoint("agent.completed", {
+                            "agent.name": self.name,
+                            "agent.iterations": self.max_iterations,
+                            "agent.tool_calls_count": len(all_tool_calls),
+                            "agent.max_iterations_reached": True,
+                            "output_length": len(final_output),
+                        })
+
+                    return AgentResult(
+                        output=final_output,
+                        tool_calls=all_tool_calls,
+                        context=context,
+                    )
+            except Exception as e:
+                # Emit error checkpoint for observability
+                if workflow_ctx:
+                    workflow_ctx._send_checkpoint("agent.failed", {
+                        "agent.name": self.name,
+                        "error": str(e),
+                        "error_type": type(e).__name__,
+                    })
+                raise
+        finally:
+            # Always reset context to prevent leakage between agent executions
+            from .context import _current_context
+            _current_context.reset(token)
+
+    async def resume_from_hitl(
+        self,
+        context: Context,
+        agent_context: Dict,
+        user_response: str,
+    ) -> AgentResult:
+        """
+        Resume agent execution after HITL pause.
+
+        This method reconstructs agent state from the checkpoint and injects
+        the user's response as the successful tool result, then continues
+        the conversation loop.
+
+        Args:
+            context: Current execution context (workflow or agent)
+            agent_context: Agent state from WaitingForUserInputException.agent_context
+            user_response: User's answer to the HITL question
+
+        Returns:
+            AgentResult with final output and tool calls
+        """
+        self.logger.info(f"Resuming agent '{self.name}' from HITL pause")
+
+        # 1. Restore conversation state
+        messages = [
+            Message(role=lm.MessageRole(msg["role"]), content=msg["content"])
+            for msg in agent_context["messages"]
+        ]
+        iteration = agent_context["iteration"]
+        all_tool_calls = agent_context["all_tool_calls"]
+
+        # 2. Restore partial tool results for current iteration
+        tool_results = agent_context["tool_results"]
+
+        # 3. Inject user response as successful tool result
+        pending_tool = agent_context["pending_tool_call"]
+        tool_results.append({
+            "tool": pending_tool["name"],
+            "result": json.dumps(user_response),
+            "error": None,
+        })
+
+        self.logger.debug(
+            f"Injected user response for tool '{pending_tool['name']}': {user_response}"
+        )
+
+        # 4. Add tool results to conversation
+        results_text = "\n".join([
+            f"Tool: {tr['tool']}\nResult: {tr['result']}"
+            if tr["error"] is None
+            else f"Tool: {tr['tool']}\nError: {tr['error']}"
+            for tr in tool_results
+        ])
+        messages.append(Message.user(
+            f"Tool results:\n{results_text}\n\n"
+            f"Please provide your final answer based on these results."
+        ))
+
+        # 5. Continue agent execution loop from next iteration
+        return await self._continue_execution_from_iteration(
+            context=context,
+            messages=messages,
+            iteration=iteration + 1,  # Next iteration
+            all_tool_calls=all_tool_calls,
+        )
+
+    async def _continue_execution_from_iteration(
+        self,
+        context: Context,
+        messages: List[Message],
+        iteration: int,
+        all_tool_calls: List[Dict],
+    ) -> AgentResult:
+        """
+        Continue agent execution from a specific iteration.
+
+        This is the core execution loop extracted to support both:
+        1. Normal execution (starting from iteration 0)
+        2. Resume after HITL (starting from iteration N)
+
+        Args:
+            context: Execution context
+            messages: Conversation history
+            iteration: Starting iteration number
+            all_tool_calls: Accumulated tool calls
+
+        Returns:
+            AgentResult with output and tool calls
+        """
+        # Extract workflow context for checkpointing
+        workflow_ctx = None
+        if hasattr(context, "_workflow_entity"):
+            workflow_ctx = context
+        elif hasattr(context, "_agent_data") and "_workflow_ctx" in context._agent_data:
+            workflow_ctx = context._agent_data["_workflow_ctx"]
+
+        # Prepare tool definitions
+        tool_defs = [
+            ToolDefinition(
+                name=name,
+                description=tool.description or f"Tool: {name}",
+                parameters=tool.input_schema if hasattr(tool, "input_schema") else {},
+            )
+            for name, tool in self.tools.items()
+        ]
+
+        # Main iteration loop (continue from specified iteration)
+        while iteration < self.max_iterations:
+            self.logger.debug(f"Agent iteration {iteration + 1}/{self.max_iterations}")
+
+            # Call LLM for next response
+            if self._language_model:
+                # Legacy API: model is a LanguageModel instance
+                request = GenerateRequest(
+                    system_prompt=self.instructions,
+                    messages=messages,
+                    tools=tool_defs if tool_defs else [],
+                )
+                request.config.temperature = self.temperature
+                if self.max_tokens:
+                    request.config.max_tokens = self.max_tokens
+                if self.top_p:
+                    request.config.top_p = self.top_p
+                response = await self._language_model.generate(request)
+
+                # Track cost for this LLM call
+                self._track_llm_cost(response, workflow_ctx)
+            else:
+                # New API: model is a string, create internal LM instance
+                request = GenerateRequest(
+                    model=self.model,
+                    system_prompt=self.instructions,
+                    messages=messages,
+                    tools=tool_defs if tool_defs else [],
+                )
+                request.config.temperature = self.temperature
+                if self.max_tokens:
+                    request.config.max_tokens = self.max_tokens
+                if self.top_p:
+                    request.config.top_p = self.top_p
+
+                # Create internal LM instance for generation
+                from .lm import _LanguageModel
+                provider, model_name = self.model.split('/', 1)
+                internal_lm = _LanguageModel(provider=provider.lower(), default_model=None)
+                response = await internal_lm.generate(request)
+
+                # Track cost for this LLM call
+                self._track_llm_cost(response, workflow_ctx)
+
+            # Add assistant response to messages
+            messages.append(Message.assistant(response.text))
+
+            # Check if LLM wants to use tools
+            if response.tool_calls:
+                self.logger.debug(f"Agent calling {len(response.tool_calls)} tool(s)")
+
+                # Store current conversation in context for potential handoffs
+                if not hasattr(context, '_agent_data'):
+                    context._agent_data = {}
+                context._agent_data["_current_conversation"] = messages
+
+                # Execute tool calls
+                tool_results = []
+                for tool_call in response.tool_calls:
+                    tool_name = tool_call["name"]
+                    tool_args_str = tool_call["arguments"]
+
+                    # Track tool call
+                    all_tool_calls.append({
+                        "name": tool_name,
+                        "arguments": tool_args_str,
+                        "iteration": iteration + 1,
+                    })
+
+                    # Execute tool
+                    try:
+                        # Parse arguments
+                        tool_args = json.loads(tool_args_str)
+
+                        # Get tool
+                        tool = self.tools.get(tool_name)
+                        if not tool:
+                            result_text = f"Error: Tool '{tool_name}' not found"
+                        else:
+                            # Execute tool
+                            result = await tool.invoke(context, **tool_args)
+
+                            # Check if this was a handoff
+                            if isinstance(result, dict) and result.get("_handoff"):
+                                self.logger.info(
+                                    f"Handoff detected to '{result['to_agent']}', "
+                                    f"terminating current agent"
+                                )
+                                # Save conversation before returning
+                                if isinstance(context, AgentContext):
+                                    await context.save_conversation_history(messages)
+                                # Return immediately with handoff result
+                                return AgentResult(
+                                    output=result["output"],
+                                    tool_calls=all_tool_calls + result.get("tool_calls", []),
+                                    context=context,
+                                    handoff_to=result["to_agent"],
+                                    handoff_metadata=result,
+                                )
+
+                            result_text = json.dumps(result) if result else "null"
+
+                        tool_results.append(
+                            {"tool": tool_name, "result": result_text, "error": None}
+                        )
+
+                    except WaitingForUserInputException as e:
+                        # HITL PAUSE: Capture agent state and propagate exception
+                        self.logger.info(f"Agent pausing for user input at iteration {iteration}")
+
+                        # Serialize messages to dict format
+                        messages_dict = [
+                            {"role": msg.role.value, "content": msg.content}
+                            for msg in messages
+                        ]
+
+                        # Enhance exception with agent execution context
+                        from .exceptions import WaitingForUserInputException
+                        raise WaitingForUserInputException(
+                            question=e.question,
+                            input_type=e.input_type,
+                            options=e.options,
+                            checkpoint_state=e.checkpoint_state,
+                            agent_context={
+                                "agent_name": self.name,
+                                "iteration": iteration,
+                                "messages": messages_dict,
+                                "tool_results": tool_results,
+                                "pending_tool_call": {
+                                    "name": tool_call["name"],
+                                    "arguments": tool_call["arguments"],
+                                    "tool_call_index": response.tool_calls.index(tool_call),
+                                },
+                                "all_tool_calls": all_tool_calls,
+                                "model_config": {
+                                    "model": self.model,
+                                    "temperature": self.temperature,
+                                    "max_tokens": self.max_tokens,
+                                    "top_p": self.top_p,
+                                },
+                            },
+                        ) from e
+
+                    except Exception as e:
+                        # Regular tool errors - log and continue
+                        self.logger.error(f"Tool execution error: {e}")
+                        tool_results.append(
+                            {"tool": tool_name, "result": None, "error": str(e)}
+                        )
+
+                # Add tool results to conversation
+                results_text = "\n".join([
+                    f"Tool: {tr['tool']}\nResult: {tr['result']}"
+                    if tr["error"] is None
+                    else f"Tool: {tr['tool']}\nError: {tr['error']}"
+                    for tr in tool_results
+                ])
+                messages.append(Message.user(
+                    f"Tool results:\n{results_text}\n\n"
+                    f"Please provide your final answer based on these results."
+                ))
+
+                # Continue loop for agent to process results
+
+            else:
+                # No tool calls - agent is done
+                self.logger.debug(f"Agent completed after {iteration + 1} iterations")
+                # Save conversation before returning
+                if isinstance(context, AgentContext):
+                    await context.save_conversation_history(messages)
+
+                # Emit completion checkpoint
+                if workflow_ctx:
+                    workflow_ctx._send_checkpoint("agent.completed", {
+                        "agent.name": self.name,
+                        "agent.iterations": iteration + 1,
+                        "agent.tool_calls_count": len(all_tool_calls),
+                        "output_length": len(response.text),
+                    })
+
+                return AgentResult(
+                    output=response.text,
+                    tool_calls=all_tool_calls,
+                    context=context,
+                )
+
+            iteration += 1
+
+        # Max iterations reached
+        self.logger.warning(f"Agent reached max iterations ({self.max_iterations})")
+        final_output = messages[-1].content if messages else "No output generated"
+        # Save conversation before returning
+        if isinstance(context, AgentContext):
+            await context.save_conversation_history(messages)
+
+        # Emit completion checkpoint with max iterations flag
+        if workflow_ctx:
+            workflow_ctx._send_checkpoint("agent.completed", {
+                "agent.name": self.name,
+                "agent.iterations": self.max_iterations,
+                "agent.tool_calls_count": len(all_tool_calls),
+                "agent.max_iterations_reached": True,
+                "output_length": len(final_output),
+            })
+
+        return AgentResult(
+            output=final_output,
+            tool_calls=all_tool_calls,
+            context=context,
+        )
+
+
+def agent(
+    _func: Optional[Callable] = None,
+    *,
+    name: Optional[str] = None,
+    model: Optional[LanguageModel] = None,
+    instructions: Optional[str] = None,
+    tools: Optional[List[Any]] = None,
+    model_name: str = "gpt-4o-mini",
+    temperature: float = 0.7,
+    max_iterations: int = 10,
+) -> Callable:
+    """
+    Decorator to register a function as an agent and automatically register it.
+
+    This decorator allows you to define agents as functions that create and return Agent instances.
+    The agent will be automatically registered in the AgentRegistry for discovery by the worker.
+
+    Args:
+        name: Agent name (defaults to function name)
+        model: Language model instance (required if not provided in function)
+        instructions: System instructions (required if not provided in function)
+        tools: List of tools available to the agent
+        model_name: Model name to use
+        temperature: LLM temperature
+        max_iterations: Maximum reasoning iterations
+
+    Returns:
+        Decorated function that returns an Agent instance
+
+    Example:
+        ```python
+        from agnt5 import agent, tool
+        from agnt5.lm import OpenAILanguageModel
+
+        @agent(
+            name="research_agent",
+            model=OpenAILanguageModel(),
+            instructions="You are a research assistant.",
+            tools=[search_web, analyze_data]
+        )
+        def create_researcher():
+            # Agent is created and registered automatically
+            pass
+
+        # Or create agent directly
+        @agent
+        def my_agent():
+            from agnt5.lm import OpenAILanguageModel
+            return Agent(
+                name="my_agent",
+                model=OpenAILanguageModel(),
+                instructions="You are a helpful assistant."
+            )
+        ```
+    """
+
+    def decorator(func: Callable) -> Callable:
+        # Determine agent name
+        agent_name = name or func.__name__
+
+        # Create the agent
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs) -> Agent:
+            # Check if function returns an Agent
+            result = func(*args, **kwargs)
+            if isinstance(result, Agent):
+                # Function creates its own agent
+                agent_instance = result
+            elif model is not None and instructions is not None:
+                # Create agent from decorator parameters
+                agent_instance = Agent(
+                    name=agent_name,
+                    model=model,
+                    instructions=instructions,
+                    tools=tools,
+                    model_name=model_name,
+                    temperature=temperature,
+                    max_iterations=max_iterations,
+                )
+            else:
+                raise ValueError(
+                    f"Agent decorator for '{agent_name}' requires either "
+                    "the decorated function to return an Agent instance, "
+                    "or 'model' and 'instructions' parameters to be provided"
+                )
+
+            # Register agent
+            AgentRegistry.register(agent_instance)
+            return agent_instance
+
+        # Create agent immediately and store reference
+        agent_instance = wrapper()
+
+        # Return the agent instance itself (so it can be used directly)
+        return agent_instance
+
+    if _func is None:
+        return decorator
+    return decorator(_func)