pydantic-ai-slim 0.0.12__py3-none-any.whl → 0.0.14__py3-none-any.whl

pydantic_ai/agent.py

@@ -7,7 +7,7 @@ from collections.abc import AsyncIterator, Awaitable, Iterator, Sequence
 from contextlib import asynccontextmanager, contextmanager
 from dataclasses import dataclass, field
 from types import FrameType
-from typing import Any, Callable, Generic, cast, final, overload
+from typing import Any, Callable, Generic, Literal, cast, final, overload
 
 import logfire_api
 from typing_extensions import assert_never
@@ -22,6 +22,7 @@ from . import (
     result,
 )
 from .result import ResultData
+from .settings import ModelSettings, UsageLimits, merge_model_settings
 from .tools import (
     AgentDeps,
     RunContext,
@@ -39,6 +40,12 @@ __all__ = ('Agent',)
 _logfire = logfire_api.Logfire(otel_scope='pydantic-ai')
 
 NoneType = type(None)
+EndStrategy = Literal['early', 'exhaustive']
+"""The strategy for handling multiple tool calls when a final result is found.
+
+- `'early'`: Stop processing other tool calls once a final result is found
+- `'exhaustive'`: Process all tool calls even after finding a final result
+"""
 
 
 @final
@@ -53,7 +60,7 @@ class Agent(Generic[AgentDeps, ResultData]):
 
     Minimal usage example:
 
-    ```py
+    ```python
     from pydantic_ai import Agent
 
     agent = Agent('openai:gpt-4o')
@@ -63,14 +70,31 @@ class Agent(Generic[AgentDeps, ResultData]):
     ```
     """
 
-    # dataclass fields mostly for my sanity — knowing what attributes are available
+    # we use dataclass fields in order to conveniently know what attributes are available
     model: models.Model | models.KnownModelName | None
     """The default model configured for this agent."""
+
     name: str | None
     """The name of the agent, used for logging.
 
     If `None`, we try to infer the agent name from the call frame when the agent is first run.
     """
+    end_strategy: EndStrategy
+    """Strategy for handling tool calls when a final result is found."""
+
+    model_settings: ModelSettings | None
+    """Optional model request settings to use for this agents's runs, by default.
+
+    Note, if `model_settings` is provided by `run`, `run_sync`, or `run_stream`, those settings will
+    be merged with this value, with the runtime argument taking priority.
+    """
+
+    last_run_messages: list[_messages.ModelMessage] | None
+    """The messages from the last run, useful when a run raised an exception.
+
+    Note: these are not used by the agent, e.g. in future runs, they are just stored for developers' convenience.
+    """
+
     _result_schema: _result.ResultSchema[ResultData] | None = field(repr=False)
     _result_validators: list[_result.ResultValidator[AgentDeps, ResultData]] = field(repr=False)
     _allow_text_result: bool = field(repr=False)
@@ -80,14 +104,8 @@ class Agent(Generic[AgentDeps, ResultData]):
     _system_prompt_functions: list[_system_prompt.SystemPromptRunner[AgentDeps]] = field(repr=False)
     _deps_type: type[AgentDeps] = field(repr=False)
     _max_result_retries: int = field(repr=False)
-    _current_result_retry: int = field(repr=False)
     _override_deps: _utils.Option[AgentDeps] = field(default=None, repr=False)
     _override_model: _utils.Option[models.Model] = field(default=None, repr=False)
-    last_run_messages: list[_messages.Message] | None = None
-    """The messages from the last run, useful when a run raised an exception.
-
-    Note: these are not used by the agent, e.g. in future runs, they are just stored for developers' convenience.
-    """
 
     def __init__(
         self,
@@ -97,18 +115,20 @@ class Agent(Generic[AgentDeps, ResultData]):
         system_prompt: str | Sequence[str] = (),
         deps_type: type[AgentDeps] = NoneType,
         name: str | None = None,
+        model_settings: ModelSettings | None = None,
         retries: int = 1,
         result_tool_name: str = 'final_result',
         result_tool_description: str | None = None,
         result_retries: int | None = None,
         tools: Sequence[Tool[AgentDeps] | ToolFuncEither[AgentDeps, ...]] = (),
         defer_model_check: bool = False,
+        end_strategy: EndStrategy = 'early',
     ):
         """Create an agent.
 
         Args:
             model: The default model to use for this agent, if not provide,
-                you must provide the model when calling the agent.
+                you must provide the model when calling it.
             result_type: The type of the result data, used to validate the result data, defaults to `str`.
             system_prompt: Static system prompts to use for this agent, you can also register system
                 prompts via a function with [`system_prompt`][pydantic_ai.Agent.system_prompt].
@@ -118,6 +138,7 @@ class Agent(Generic[AgentDeps, ResultData]):
                 or add a type hint `: Agent[None, <return type>]`.
             name: The name of the agent, used for logging. If `None`, we try to infer the agent name from the call frame
                 when the agent is first run.
+            model_settings: Optional model request settings to use for this agent's runs, by default.
             retries: The default number of retries to allow before raising an error.
             result_tool_name: The name of the tool to use for the final result.
             result_tool_description: The description of the final result tool.
@@ -129,13 +150,18 @@ class Agent(Generic[AgentDeps, ResultData]):
                 which checks for the necessary environment variables. Set this to `false`
                 to defer the evaluation until the first run. Useful if you want to
                 [override the model][pydantic_ai.Agent.override] for testing.
+            end_strategy: Strategy for handling tool calls that are requested alongside a final result.
+                See [`EndStrategy`][pydantic_ai.agent.EndStrategy] for more information.
         """
         if model is None or defer_model_check:
             self.model = model
         else:
             self.model = models.infer_model(model)
 
+        self.end_strategy = end_strategy
         self.name = name
+        self.model_settings = model_settings
+        self.last_run_messages = None
         self._result_schema = _result.ResultSchema[result_type].build(
             result_type, result_tool_name, result_tool_description
         )
@@ -153,25 +179,39 @@ class Agent(Generic[AgentDeps, ResultData]):
         self._deps_type = deps_type
         self._system_prompt_functions = []
         self._max_result_retries = result_retries if result_retries is not None else retries
-        self._current_result_retry = 0
         self._result_validators = []
 
     async def run(
         self,
         user_prompt: str,
         *,
-        message_history: list[_messages.Message] | None = None,
+        message_history: list[_messages.ModelMessage] | None = None,
         model: models.Model | models.KnownModelName | None = None,
         deps: AgentDeps = None,
+        model_settings: ModelSettings | None = None,
+        usage_limits: UsageLimits | None = None,
         infer_name: bool = True,
     ) -> result.RunResult[ResultData]:
         """Run the agent with a user prompt in async mode.
 
+        Example:
+        ```python
+        from pydantic_ai import Agent
+
+        agent = Agent('openai:gpt-4o')
+
+        result_sync = agent.run_sync('What is the capital of Italy?')
+        print(result_sync.data)
+        #> Rome
+        ```
+
         Args:
             user_prompt: User input to start/continue the conversation.
             message_history: History of the conversation so far.
             model: Optional model to use for this run, required if `model` was not set when creating the agent.
             deps: Optional dependencies to use for this run.
+            model_settings: Optional settings to use for this model's request.
+            usage_limits: Optional limits on model request count or token usage.
             infer_name: Whether to try to infer the agent name from the call frame if it's not set.
 
         Returns:
@@ -182,6 +222,7 @@ class Agent(Generic[AgentDeps, ResultData]):
         model_used, mode_selection = await self._get_model(model)
 
         deps = self._get_deps(deps)
+        new_message_index = len(message_history) if message_history else 0
 
         with _logfire.span(
             '{agent_name} run {prompt=}',
@@ -191,67 +232,91 @@ class Agent(Generic[AgentDeps, ResultData]):
             model_name=model_used.name(),
             agent_name=self.name or 'agent',
         ) as run_span:
-            new_message_index, messages = await self._prepare_messages(deps, user_prompt, message_history)
-            self.last_run_messages = messages
+            run_context = RunContext(deps, 0, [], None, model_used)
+            messages = await self._prepare_messages(user_prompt, message_history, run_context)
+            self.last_run_messages = run_context.messages = messages
 
             for tool in self._function_tools.values():
                 tool.current_retry = 0
 
-            cost = result.Cost()
+            usage = result.Usage(requests=0)
+            model_settings = merge_model_settings(self.model_settings, model_settings)
+            usage_limits = usage_limits or UsageLimits()
 
             run_step = 0
             while True:
+                usage_limits.check_before_request(usage)
+
                 run_step += 1
                 with _logfire.span('preparing model and tools {run_step=}', run_step=run_step):
-                    agent_model = await self._prepare_model(model_used, deps)
+                    agent_model = await self._prepare_model(run_context)
 
                 with _logfire.span('model request', run_step=run_step) as model_req_span:
-                    model_response, request_cost = await agent_model.request(messages)
+                    model_response, request_usage = await agent_model.request(messages, model_settings)
                     model_req_span.set_attribute('response', model_response)
-                    model_req_span.set_attribute('cost', request_cost)
-                    model_req_span.message = f'model request -> {model_response.role}'
+                    model_req_span.set_attribute('usage', request_usage)
 
                 messages.append(model_response)
-                cost += request_cost
+                usage += request_usage
+                usage.requests += 1
+                usage_limits.check_tokens(request_usage)
 
                 with _logfire.span('handle model response', run_step=run_step) as handle_span:
-                    final_result, response_messages = await self._handle_model_response(model_response, deps)
+                    final_result, tool_responses = await self._handle_model_response(model_response, run_context)
 
-                    # Add all messages to the conversation
-                    messages.extend(response_messages)
+                    if tool_responses:
+                        # Add parts to the conversation as a new message
+                        messages.append(_messages.ModelRequest(tool_responses))
 
                     # Check if we got a final result
                     if final_result is not None:
                         result_data = final_result.data
                         run_span.set_attribute('all_messages', messages)
-                        run_span.set_attribute('cost', cost)
+                        run_span.set_attribute('usage', usage)
                         handle_span.set_attribute('result', result_data)
                         handle_span.message = 'handle model response -> final result'
-                        return result.RunResult(messages, new_message_index, result_data, cost)
+                        return result.RunResult(messages, new_message_index, result_data, usage)
                     else:
                         # continue the conversation
-                        handle_span.set_attribute('tool_responses', response_messages)
-                        response_msgs = ' '.join(r.role for r in response_messages)
-                        handle_span.message = f'handle model response -> {response_msgs}'
+                        handle_span.set_attribute('tool_responses', tool_responses)
+                        tool_responses_str = ' '.join(r.part_kind for r in tool_responses)
+                        handle_span.message = f'handle model response -> {tool_responses_str}'
 
     def run_sync(
         self,
         user_prompt: str,
         *,
-        message_history: list[_messages.Message] | None = None,
+        message_history: list[_messages.ModelMessage] | None = None,
         model: models.Model | models.KnownModelName | None = None,
         deps: AgentDeps = None,
+        model_settings: ModelSettings | None = None,
+        usage_limits: UsageLimits | None = None,
         infer_name: bool = True,
     ) -> result.RunResult[ResultData]:
         """Run the agent with a user prompt synchronously.
 
-        This is a convenience method that wraps `self.run` with `loop.run_until_complete()`.
+        This is a convenience method that wraps [`self.run`][pydantic_ai.Agent.run] with `loop.run_until_complete(...)`.
+        You therefore can't use this method inside async code or if there's an active event loop.
+
+        Example:
+        ```python
+        from pydantic_ai import Agent
+
+        agent = Agent('openai:gpt-4o')
+
+        async def main():
+            result = await agent.run('What is the capital of France?')
+            print(result.data)
+            #> Paris
+        ```
 
         Args:
             user_prompt: User input to start/continue the conversation.
             message_history: History of the conversation so far.
             model: Optional model to use for this run, required if `model` was not set when creating the agent.
             deps: Optional dependencies to use for this run.
+            model_settings: Optional settings to use for this model's request.
+            usage_limits: Optional limits on model request count or token usage.
             infer_name: Whether to try to infer the agent name from the call frame if it's not set.
 
         Returns:
@@ -259,9 +324,16 @@ class Agent(Generic[AgentDeps, ResultData]):
         """
         if infer_name and self.name is None:
             self._infer_name(inspect.currentframe())
-        loop = asyncio.get_event_loop()
-        return loop.run_until_complete(
-            self.run(user_prompt, message_history=message_history, model=model, deps=deps, infer_name=False)
+        return asyncio.get_event_loop().run_until_complete(
+            self.run(
+                user_prompt,
+                message_history=message_history,
+                model=model,
+                deps=deps,
+                model_settings=model_settings,
+                usage_limits=usage_limits,
+                infer_name=False,
+            )
         )
 
     @asynccontextmanager
@@ -269,18 +341,34 @@ class Agent(Generic[AgentDeps, ResultData]):
         self,
         user_prompt: str,
         *,
-        message_history: list[_messages.Message] | None = None,
+        message_history: list[_messages.ModelMessage] | None = None,
         model: models.Model | models.KnownModelName | None = None,
         deps: AgentDeps = None,
+        model_settings: ModelSettings | None = None,
+        usage_limits: UsageLimits | None = None,
         infer_name: bool = True,
     ) -> AsyncIterator[result.StreamedRunResult[AgentDeps, ResultData]]:
         """Run the agent with a user prompt in async mode, returning a streamed response.
 
+        Example:
+        ```python
+        from pydantic_ai import Agent
+
+        agent = Agent('openai:gpt-4o')
+
+        async def main():
+            async with agent.run_stream('What is the capital of the UK?') as response:
+                print(await response.get_data())
+                #> London
+        ```
+
         Args:
             user_prompt: User input to start/continue the conversation.
             message_history: History of the conversation so far.
             model: Optional model to use for this run, required if `model` was not set when creating the agent.
             deps: Optional dependencies to use for this run.
+            model_settings: Optional settings to use for this model's request.
+            usage_limits: Optional limits on model request count or token usage.
             infer_name: Whether to try to infer the agent name from the call frame if it's not set.
 
         Returns:
@@ -293,6 +381,7 @@ class Agent(Generic[AgentDeps, ResultData]):
         model_used, mode_selection = await self._get_model(model)
 
         deps = self._get_deps(deps)
+        new_message_index = len(message_history) if message_history else 0
 
         with _logfire.span(
             '{agent_name} run stream {prompt=}',
@@ -302,60 +391,89 @@ class Agent(Generic[AgentDeps, ResultData]):
             model_name=model_used.name(),
             agent_name=self.name or 'agent',
         ) as run_span:
-            new_message_index, messages = await self._prepare_messages(deps, user_prompt, message_history)
-            self.last_run_messages = messages
+            run_context = RunContext(deps, 0, [], None, model_used)
+            messages = await self._prepare_messages(user_prompt, message_history, run_context)
+            self.last_run_messages = run_context.messages = messages
 
             for tool in self._function_tools.values():
                 tool.current_retry = 0
 
-            cost = result.Cost()
+            usage = result.Usage()
+            model_settings = merge_model_settings(self.model_settings, model_settings)
+            usage_limits = usage_limits or UsageLimits()
 
             run_step = 0
             while True:
                 run_step += 1
+                usage_limits.check_before_request(usage)
 
                 with _logfire.span('preparing model and tools {run_step=}', run_step=run_step):
-                    agent_model = await self._prepare_model(model_used, deps)
+                    agent_model = await self._prepare_model(run_context)
 
                 with _logfire.span('model request {run_step=}', run_step=run_step) as model_req_span:
-                    async with agent_model.request_stream(messages) as model_response:
+                    async with agent_model.request_stream(messages, model_settings) as model_response:
+                        usage.requests += 1
                         model_req_span.set_attribute('response_type', model_response.__class__.__name__)
                         # We want to end the "model request" span here, but we can't exit the context manager
                         # in the traditional way
                         model_req_span.__exit__(None, None, None)
 
                         with _logfire.span('handle model response') as handle_span:
-                            final_result, response_messages = await self._handle_streamed_model_response(
-                                model_response, deps
-                            )
-
-                            # Add all messages to the conversation
-                            messages.extend(response_messages)
+                            maybe_final_result = await self._handle_streamed_model_response(model_response, run_context)
 
                             # Check if we got a final result
-                            if final_result is not None:
-                                result_stream = final_result.data
-                                run_span.set_attribute('all_messages', messages)
-                                handle_span.set_attribute('result_type', result_stream.__class__.__name__)
+                            if isinstance(maybe_final_result, _MarkFinalResult):
+                                result_stream = maybe_final_result.data
+                                result_tool_name = maybe_final_result.tool_name
                                 handle_span.message = 'handle model response -> final result'
+
+                                async def on_complete():
+                                    """Called when the stream has completed.
+
+                                    The model response will have been added to messages by now
+                                    by `StreamedRunResult._marked_completed`.
+                                    """
+                                    last_message = messages[-1]
+                                    assert isinstance(last_message, _messages.ModelResponse)
+                                    tool_calls = [
+                                        part for part in last_message.parts if isinstance(part, _messages.ToolCallPart)
+                                    ]
+                                    parts = await self._process_function_tools(
+                                        tool_calls, result_tool_name, run_context
+                                    )
+                                    if parts:
+                                        messages.append(_messages.ModelRequest(parts))
+                                    run_span.set_attribute('all_messages', messages)
+
                                 yield result.StreamedRunResult(
                                     messages,
                                     new_message_index,
-                                    cost,
+                                    usage,
+                                    usage_limits,
                                     result_stream,
                                     self._result_schema,
-                                    deps,
+                                    run_context,
                                     self._result_validators,
-                                    lambda m: run_span.set_attribute('all_messages', messages),
+                                    result_tool_name,
+                                    on_complete,
                                 )
                                 return
                             else:
                                 # continue the conversation
-                                handle_span.set_attribute('tool_responses', response_messages)
-                                response_msgs = ' '.join(r.role for r in response_messages)
-                                handle_span.message = f'handle model response -> {response_msgs}'
-                                # the model_response should have been fully streamed by now, we can add it's cost
-                                cost += model_response.cost()
+                                model_response_msg, tool_responses = maybe_final_result
+                                # if we got a model response add that to messages
+                                messages.append(model_response_msg)
+                                if tool_responses:
+                                    # if we got one or more tool response parts, add a model request message
+                                    messages.append(_messages.ModelRequest(tool_responses))
+
+                                handle_span.set_attribute('tool_responses', tool_responses)
+                                tool_responses_str = ' '.join(r.part_kind for r in tool_responses)
+                                handle_span.message = f'handle model response -> {tool_responses_str}'
+                                # the model_response should have been fully streamed by now, we can add its usage
+                                model_response_usage = model_response.usage()
+                                usage += model_response_usage
+                                usage_limits.check_tokens(usage)
 
     @contextmanager
     def override(
@@ -367,6 +485,7 @@ class Agent(Generic[AgentDeps, ResultData]):
         """Context manager to temporarily override agent dependencies and model.
 
         This is particularly useful when testing.
+        You can find an example of this [here](../testing-evals.md#overriding-model-via-pytest-fixtures).
 
         Args:
             deps: The dependencies to use instead of the dependencies passed to the agent run.
@@ -415,14 +534,14 @@ class Agent(Generic[AgentDeps, ResultData]):
     ) -> _system_prompt.SystemPromptFunc[AgentDeps]:
         """Decorator to register a system prompt function.
 
-        Optionally takes [`RunContext`][pydantic_ai.tools.RunContext] as it's only argument.
+        Optionally takes [`RunContext`][pydantic_ai.tools.RunContext] as its only argument.
         Can decorate a sync or async functions.
 
         Overloads for every possible signature of `system_prompt` are included so the decorator doesn't obscure
         the type of the function, see `tests/typed_agent.py` for tests.
 
         Example:
-        ```py
+        ```python
         from pydantic_ai import Agent, RunContext
 
         agent = Agent('test', deps_type=str)
@@ -466,14 +585,14 @@ class Agent(Generic[AgentDeps, ResultData]):
     ) -> _result.ResultValidatorFunc[AgentDeps, ResultData]:
         """Decorator to register a result validator function.
 
-        Optionally takes [`RunContext`][pydantic_ai.tools.RunContext] as it's first argument.
+        Optionally takes [`RunContext`][pydantic_ai.tools.RunContext] as its first argument.
         Can decorate a sync or async functions.
 
         Overloads for every possible signature of `result_validator` are included so the decorator doesn't obscure
         the type of the function, see `tests/typed_agent.py` for tests.
 
         Example:
-        ```py
+        ```python
         from pydantic_ai import Agent, ModelRetry, RunContext
 
         agent = Agent('test', deps_type=str)
@@ -523,13 +642,13 @@ class Agent(Generic[AgentDeps, ResultData]):
         Can decorate a sync or async functions.
 
         The docstring is inspected to extract both the tool description and description of each parameter,
-        [learn more](../agents.md#function-tools-and-schema).
+        [learn more](../tools.md#function-tools-and-schema).
 
         We can't add overloads for every possible signature of tool, since the return type is a recursive union
         so the signature of functions decorated with `@agent.tool` is obscured.
 
         Example:
-        ```py
+        ```python
         from pydantic_ai import Agent, RunContext
 
         agent = Agent('test', deps_type=int)
@@ -595,13 +714,13 @@ class Agent(Generic[AgentDeps, ResultData]):
         Can decorate a sync or async functions.
 
         The docstring is inspected to extract both the tool description and description of each parameter,
-        [learn more](../agents.md#function-tools-and-schema).
+        [learn more](../tools.md#function-tools-and-schema).
 
         We can't add overloads for every possible signature of tool, since the return type is a recursive union
         so the signature of functions decorated with `@agent.tool` is obscured.
 
         Example:
-        ```py
+        ```python
         from pydantic_ai import Agent, RunContext
 
         agent = Agent('test')
@@ -696,193 +815,266 @@ class Agent(Generic[AgentDeps, ResultData]):
 
         return model_, mode_selection
 
-    async def _prepare_model(self, model: models.Model, deps: AgentDeps) -> models.AgentModel:
-        """Create building tools and create an agent model."""
+    async def _prepare_model(self, run_context: RunContext[AgentDeps]) -> models.AgentModel:
+        """Build tools and create an agent model."""
         function_tools: list[ToolDefinition] = []
 
         async def add_tool(tool: Tool[AgentDeps]) -> None:
-            ctx = RunContext(deps, tool.current_retry, tool.name)
+            ctx = run_context.replace_with(retry=tool.current_retry, tool_name=tool.name)
             if tool_def := await tool.prepare_tool_def(ctx):
                 function_tools.append(tool_def)
 
         await asyncio.gather(*map(add_tool, self._function_tools.values()))
 
-        return await model.agent_model(
+        return await run_context.model.agent_model(
             function_tools=function_tools,
             allow_text_result=self._allow_text_result,
             result_tools=self._result_schema.tool_defs() if self._result_schema is not None else [],
         )
 
     async def _prepare_messages(
-        self, deps: AgentDeps, user_prompt: str, message_history: list[_messages.Message] | None
-    ) -> tuple[int, list[_messages.Message]]:
-        # if message history includes system prompts, we don't want to regenerate them
-        if message_history and any(m.role == 'system' for m in message_history):
+        self, user_prompt: str, message_history: list[_messages.ModelMessage] | None, run_context: RunContext[AgentDeps]
+    ) -> list[_messages.ModelMessage]:
+        if message_history:
             # shallow copy messages
             messages = message_history.copy()
+            messages.append(_messages.ModelRequest([_messages.UserPromptPart(user_prompt)]))
         else:
-            messages = await self._init_messages(deps)
-            if message_history:
-                messages += message_history
+            parts = await self._sys_parts(run_context)
+            parts.append(_messages.UserPromptPart(user_prompt))
+            messages: list[_messages.ModelMessage] = [_messages.ModelRequest(parts)]
 
-        new_message_index = len(messages)
-        messages.append(_messages.UserPrompt(user_prompt))
-        return new_message_index, messages
+        return messages
 
     async def _handle_model_response(
-        self, model_response: _messages.ModelAnyResponse, deps: AgentDeps
-    ) -> tuple[_MarkFinalResult[ResultData] | None, list[_messages.Message]]:
+        self, model_response: _messages.ModelResponse, run_context: RunContext[AgentDeps]
+    ) -> tuple[_MarkFinalResult[ResultData] | None, list[_messages.ModelRequestPart]]:
         """Process a non-streamed response from the model.
 
         Returns:
-            A tuple of `(final_result, messages)`. If `final_result` is not `None`, the conversation should end.
+            A tuple of `(final_result, request parts)`. If `final_result` is not `None`, the conversation should end.
         """
-        if model_response.role == 'model-text-response':
-            # plain string response
-            if self._allow_text_result:
-                result_data_input = cast(ResultData, model_response.content)
+        texts: list[str] = []
+        tool_calls: list[_messages.ToolCallPart] = []
+        for part in model_response.parts:
+            if isinstance(part, _messages.TextPart):
+                # ignore empty content for text parts, see #437
+                if part.content:
+                    texts.append(part.content)
+            else:
+                tool_calls.append(part)
+
+        if texts:
+            text = '\n\n'.join(texts)
+            return await self._handle_text_response(text, run_context)
+        elif tool_calls:
+            return await self._handle_structured_response(tool_calls, run_context)
+        else:
+            raise exceptions.UnexpectedModelBehavior('Received empty model response')
+
+    async def _handle_text_response(
+        self, text: str, run_context: RunContext[AgentDeps]
+    ) -> tuple[_MarkFinalResult[ResultData] | None, list[_messages.ModelRequestPart]]:
+        """Handle a plain text response from the model for non-streaming responses."""
+        if self._allow_text_result:
+            result_data_input = cast(ResultData, text)
+            try:
+                result_data = await self._validate_result(result_data_input, run_context, None)
+            except _result.ToolRetryError as e:
+                self._incr_result_retry(run_context)
+                return None, [e.tool_retry]
+            else:
+                return _MarkFinalResult(result_data, None), []
+        else:
+            self._incr_result_retry(run_context)
+            response = _messages.RetryPromptPart(
+                content='Plain text responses are not permitted, please call one of the functions instead.',
+            )
+            return None, [response]
+
+    async def _handle_structured_response(
+        self, tool_calls: list[_messages.ToolCallPart], run_context: RunContext[AgentDeps]
+    ) -> tuple[_MarkFinalResult[ResultData] | None, list[_messages.ModelRequestPart]]:
+        """Handle a structured response containing tool calls from the model for non-streaming responses."""
+        assert tool_calls, 'Expected at least one tool call'
+
+        # first look for the result tool call
+        final_result: _MarkFinalResult[ResultData] | None = None
+
+        parts: list[_messages.ModelRequestPart] = []
+        if result_schema := self._result_schema:
+            if match := result_schema.find_tool(tool_calls):
+                call, result_tool = match
                 try:
-                    result_data = await self._validate_result(result_data_input, deps, None)
+                    result_data = result_tool.validate(call)
+                    result_data = await self._validate_result(result_data, run_context, call)
                 except _result.ToolRetryError as e:
-                    self._incr_result_retry()
-                    return None, [e.tool_retry]
+                    self._incr_result_retry(run_context)
+                    parts.append(e.tool_retry)
                 else:
-                    return _MarkFinalResult(result_data), []
-            else:
-                self._incr_result_retry()
-                response = _messages.RetryPrompt(
-                    content='Plain text responses are not permitted, please call one of the functions instead.',
+                    final_result = _MarkFinalResult(result_data, call.tool_name)
+
+        # Then build the other request parts based on end strategy
+        parts += await self._process_function_tools(tool_calls, final_result and final_result.tool_name, run_context)
+
+        return final_result, parts
+
+    async def _process_function_tools(
+        self,
+        tool_calls: list[_messages.ToolCallPart],
+        result_tool_name: str | None,
+        run_context: RunContext[AgentDeps],
+    ) -> list[_messages.ModelRequestPart]:
+        """Process function (non-result) tool calls in parallel.
+
+        Also add stub return parts for any other tools that need it.
+        """
+        parts: list[_messages.ModelRequestPart] = []
+        tasks: list[asyncio.Task[_messages.ModelRequestPart]] = []
+
+        stub_function_tools = bool(result_tool_name) and self.end_strategy == 'early'
+
+        # we rely on the fact that if we found a result, it's the first result tool in the last
+        found_used_result_tool = False
+        for call in tool_calls:
+            if call.tool_name == result_tool_name and not found_used_result_tool:
+                found_used_result_tool = True
+                parts.append(
+                    _messages.ToolReturnPart(
+                        tool_name=call.tool_name,
+                        content='Final result processed.',
+                        tool_call_id=call.tool_call_id,
+                    )
                 )
-                return None, [response]
-        elif model_response.role == 'model-structured-response':
-            if self._result_schema is not None:
-                # if there's a result schema, and any of the calls match one of its tools, return the result
-                # NOTE: this means we ignore any other tools called here
-                if match := self._result_schema.find_tool(model_response):
-                    call, result_tool = match
-                    try:
-                        result_data = result_tool.validate(call)
-                        result_data = await self._validate_result(result_data, deps, call)
-                    except _result.ToolRetryError as e:
-                        self._incr_result_retry()
-                        return None, [e.tool_retry]
-                    else:
-                        # Add a ToolReturn message for the schema tool call
-                        tool_return = _messages.ToolReturn(
+            elif tool := self._function_tools.get(call.tool_name):
+                if stub_function_tools:
+                    parts.append(
+                        _messages.ToolReturnPart(
                             tool_name=call.tool_name,
-                            content='Final result processed.',
-                            tool_id=call.tool_id,
+                            content='Tool not executed - a final result was already processed.',
+                            tool_call_id=call.tool_call_id,
                         )
-                        return _MarkFinalResult(result_data), [tool_return]
-
-            if not model_response.calls:
-                raise exceptions.UnexpectedModelBehavior('Received empty tool call message')
-
-            # otherwise we run all tool functions in parallel
-            messages: list[_messages.Message] = []
-            tasks: list[asyncio.Task[_messages.Message]] = []
-            for call in model_response.calls:
-                if tool := self._function_tools.get(call.tool_name):
-                    tasks.append(asyncio.create_task(tool.run(deps, call), name=call.tool_name))
+                    )
                 else:
-                    messages.append(self._unknown_tool(call.tool_name))
+                    tasks.append(asyncio.create_task(tool.run(call, run_context), name=call.tool_name))
+            elif self._result_schema is not None and call.tool_name in self._result_schema.tools:
+                # if tool_name is in _result_schema, it means we found a result tool but an error occurred in
+                # validation, we don't add another part here
+                if result_tool_name is not None:
+                    parts.append(
+                        _messages.ToolReturnPart(
+                            tool_name=call.tool_name,
+                            content='Result tool not used - a final result was already processed.',
+                            tool_call_id=call.tool_call_id,
+                        )
+                    )
+            else:
+                parts.append(self._unknown_tool(call.tool_name, run_context))
 
+        # Run all tool tasks in parallel
+        if tasks:
             with _logfire.span('running {tools=}', tools=[t.get_name() for t in tasks]):
-                task_results: Sequence[_messages.Message] = await asyncio.gather(*tasks)
-                messages.extend(task_results)
-            return None, messages
-        else:
-            assert_never(model_response)
+                task_results: Sequence[_messages.ModelRequestPart] = await asyncio.gather(*tasks)
+                parts.extend(task_results)
+        return parts
 
     async def _handle_streamed_model_response(
-        self, model_response: models.EitherStreamedResponse, deps: AgentDeps
-    ) -> tuple[_MarkFinalResult[models.EitherStreamedResponse] | None, list[_messages.Message]]:
+        self,
+        model_response: models.EitherStreamedResponse,
+        run_context: RunContext[AgentDeps],
+    ) -> (
+        _MarkFinalResult[models.EitherStreamedResponse]
+        | tuple[_messages.ModelResponse, list[_messages.ModelRequestPart]]
+    ):
         """Process a streamed response from the model.
 
         Returns:
-            A tuple of (final_result, messages). If final_result is not None, the conversation should end.
+            Either a final result or a tuple of the model response and the tool responses for the next request.
+            If a final result is returned, the conversation should end.
         """
         if isinstance(model_response, models.StreamTextResponse):
             # plain string response
             if self._allow_text_result:
-                return _MarkFinalResult(model_response), []
+                return _MarkFinalResult(model_response, None)
             else:
-                self._incr_result_retry()
-                response = _messages.RetryPrompt(
+                self._incr_result_retry(run_context)
+                response = _messages.RetryPromptPart(
                     content='Plain text responses are not permitted, please call one of the functions instead.',
                 )
-                # stream the response, so cost is correct
+                # stream the response, so usage is correct
                 async for _ in model_response:
                     pass
 
-                return None, [response]
-        else:
-            assert isinstance(model_response, models.StreamStructuredResponse), f'Unexpected response: {model_response}'
+                text = ''.join(model_response.get(final=True))
+                return _messages.ModelResponse([_messages.TextPart(text)]), [response]
+        elif isinstance(model_response, models.StreamStructuredResponse):
             if self._result_schema is not None:
                 # if there's a result schema, iterate over the stream until we find at least one tool
                 # NOTE: this means we ignore any other tools called here
                 structured_msg = model_response.get()
-                while not structured_msg.calls:
+                while not structured_msg.parts:
                     try:
                         await model_response.__anext__()
                     except StopAsyncIteration:
                         break
                     structured_msg = model_response.get()
 
-                if match := self._result_schema.find_tool(structured_msg):
+                if match := self._result_schema.find_tool(structured_msg.parts):
                     call, _ = match
-                    tool_return = _messages.ToolReturn(
-                        tool_name=call.tool_name,
-                        content='Final result processed.',
-                        tool_id=call.tool_id,
-                    )
-                    return _MarkFinalResult(model_response), [tool_return]
+                    return _MarkFinalResult(model_response, call.tool_name)
 
             # the model is calling a tool function, consume the response to get the next message
             async for _ in model_response:
                 pass
-            structured_msg = model_response.get()
-            if not structured_msg.calls:
+            model_response_msg = model_response.get()
+            if not model_response_msg.parts:
                 raise exceptions.UnexpectedModelBehavior('Received empty tool call message')
-            messages: list[_messages.Message] = [structured_msg]
 
             # we now run all tool functions in parallel
-            tasks: list[asyncio.Task[_messages.Message]] = []
-            for call in structured_msg.calls:
-                if tool := self._function_tools.get(call.tool_name):
-                    tasks.append(asyncio.create_task(tool.run(deps, call), name=call.tool_name))
-                else:
-                    messages.append(self._unknown_tool(call.tool_name))
+            tasks: list[asyncio.Task[_messages.ModelRequestPart]] = []
+            parts: list[_messages.ModelRequestPart] = []
+            for item in model_response_msg.parts:
+                if isinstance(item, _messages.ToolCallPart):
+                    call = item
+                    if tool := self._function_tools.get(call.tool_name):
+                        tasks.append(asyncio.create_task(tool.run(call, run_context), name=call.tool_name))
+                    else:
+                        parts.append(self._unknown_tool(call.tool_name, run_context))
 
             with _logfire.span('running {tools=}', tools=[t.get_name() for t in tasks]):
-                task_results: Sequence[_messages.Message] = await asyncio.gather(*tasks)
-                messages.extend(task_results)
-            return None, messages
+                task_results: Sequence[_messages.ModelRequestPart] = await asyncio.gather(*tasks)
+                parts.extend(task_results)
+            return model_response_msg, parts
+        else:
+            assert_never(model_response)
 
     async def _validate_result(
-        self, result_data: ResultData, deps: AgentDeps, tool_call: _messages.ToolCall | None
+        self,
+        result_data: ResultData,
+        run_context: RunContext[AgentDeps],
+        tool_call: _messages.ToolCallPart | None,
     ) -> ResultData:
         for validator in self._result_validators:
-            result_data = await validator.validate(result_data, deps, self._current_result_retry, tool_call)
+            result_data = await validator.validate(result_data, tool_call, run_context)
         return result_data
 
-    def _incr_result_retry(self) -> None:
-        self._current_result_retry += 1
-        if self._current_result_retry > self._max_result_retries:
+    def _incr_result_retry(self, run_context: RunContext[AgentDeps]) -> None:
+        run_context.retry += 1
+        if run_context.retry > self._max_result_retries:
             raise exceptions.UnexpectedModelBehavior(
                 f'Exceeded maximum retries ({self._max_result_retries}) for result validation'
             )
 
-    async def _init_messages(self, deps: AgentDeps) -> list[_messages.Message]:
+    async def _sys_parts(self, run_context: RunContext[AgentDeps]) -> list[_messages.ModelRequestPart]:
         """Build the initial messages for the conversation."""
-        messages: list[_messages.Message] = [_messages.SystemPrompt(p) for p in self._system_prompts]
+        messages: list[_messages.ModelRequestPart] = [_messages.SystemPromptPart(p) for p in self._system_prompts]
         for sys_prompt_runner in self._system_prompt_functions:
-            prompt = await sys_prompt_runner.run(deps)
-            messages.append(_messages.SystemPrompt(prompt))
+            prompt = await sys_prompt_runner.run(run_context)
+            messages.append(_messages.SystemPromptPart(prompt))
         return messages
 
-    def _unknown_tool(self, tool_name: str) -> _messages.RetryPrompt:
-        self._incr_result_retry()
+    def _unknown_tool(self, tool_name: str, run_context: RunContext[AgentDeps]) -> _messages.RetryPromptPart:
+        self._incr_result_retry(run_context)
         names = list(self._function_tools.keys())
         if self._result_schema:
             names.extend(self._result_schema.tool_names())
@@ -890,7 +1082,7 @@ class Agent(Generic[AgentDeps, ResultData]):
             msg = f'Available tools: {", ".join(names)}'
         else:
             msg = 'No tools available.'
-        return _messages.RetryPrompt(content=f'Unknown tool name: {tool_name!r}. {msg}')
+        return _messages.RetryPromptPart(content=f'Unknown tool name: {tool_name!r}. {msg}')
 
     def _get_deps(self, deps: AgentDeps) -> AgentDeps:
         """Get deps for a run.
@@ -934,3 +1126,6 @@ class _MarkFinalResult(Generic[ResultData]):
     """
 
     data: ResultData
+    """The final result data."""
+    tool_name: str | None
+    """Name of the final result tool, None if the result is a string."""