Towards #345
## Summary:
Using a `dict` or `Mapping` isn't strict-mode compliant. But we were
checking for the literal `True` whereas the value can also be an array,
for example. Fix that.
## Test Plan:
Unit tests
## Summary:
Adds tracing and tests for tracing.
- Tools are added to the agents
- Theres a span for the mcp tools lookup
- Functions have MCP data
## Test Plan:
Unit tests
.
## Summary:
Adds tracing and tests for tracing.
- Tools are added to the agents
- Theres a span for the mcp tools lookup
- Functions have MCP data
## Test Plan:
Unit tests
.
This pull request introduces functionality for visualizing agent
structures using Graphviz. The changes include adding a new dependency,
implementing functions to generate and draw graphs, and adding tests for
these functions.
New functionality for visualizing agent structures:
* Added `graphviz` as a new dependency in `pyproject.toml`.
* Implemented functions in `src/agents/visualizations.py` to generate
and draw graphs for agents using Graphviz. These functions include
`get_main_graph`, `get_all_nodes`, `get_all_edges`, and `draw_graph`.
Testing the new visualization functionality:
* Added tests in `tests/test_visualizations.py` to verify the
correctness of the graph generation and drawing functions. The tests
cover `get_main_graph`, `get_all_nodes`, `get_all_edges`, and
`draw_graph`.
For example, given the following code:
```python
from agents import Agent, function_tool
from agents.visualizations import draw_graph
@function_tool
def get_weather(city: str) -> str:
return f"The weather in {city} is sunny."
spanish_agent = Agent(
name="Spanish agent",
instructions="You only speak Spanish.",
)
english_agent = Agent(
name="English agent",
instructions="You only speak English",
)
triage_agent = Agent(
name="Triage agent",
instructions="Handoff to the appropriate agent based on the language of the request.",
handoffs=[spanish_agent, english_agent],
tools=[get_weather],
)
draw_graph(triage_agent)
```
Generates the following image:
<img width="614" alt="Screenshot 2025-03-13 at 18 36 23"
src="https://github.com/user-attachments/assets/d01fe502-6886-4efb-aaf8-c92e4524b0fe"
/>
Previously the stream was closing as soon as all the audio was added but
didn't wait for it to be finished. Additionally the audio might seem
chopped off if there is no additional silence so this PR also adds one
second of silence before exiting the program.
## Summary:
#263 added this behavior. The goal was to prevent infinite loops when
tool choice was set. The key change I'm making is:
1. Making it configurable on the agent.
2. Doing bookkeeping in the Runner to track this, to prevent mutating
agents.
3. Not resetting the global tool choice in RunConfig.
## Test Plan:
Unit tests.
.
## Summary:
#263 added this behavior. The goal was to prevent infinite loops when tool choice was set. The key change I'm making is:
1. Making it configurable on the agent.
2. Doing bookkeeping in the Runner to track this, to prevent mutating agents.
3. Not resetting the global tool choice in RunConfig.
## Test Plan:
Unit tests.
.
### Summary:
Spins up a stdio server with some local files, then asks the model
questions.
### Test Plan:
Run the example, see it work.
(repeat of #322)
### Summary:
This enables users to **use** MCP inside the SDK.
1. You add a list of MCP servers to `Agent`, via `mcp_server=[...]`
2. When an agent runs, we look up its MCP tools and add them to the list
of tools.
3. When a tool call occurs, we call the relevant MCP server.
Notes:
1. There's some refactoring to make sure we send the full list of tools
to the Runner/Model etc.
2. Right now, you could have a locally defined tool that conflicts with
an MCP defined tool. I didn't add errors for that, will do in a
followup.
### Test Plan:
See unit tests. Also has an end to end example next PR.
---
[//]: # (BEGIN SAPLING FOOTER)
* #324
* #322
* __->__ #321
* #320
### Summary:
1. Add the MCP dep for python 3.10, since it doesn't support 3.9 and
below
2. Create MCPServer, which is the agents SDK representation of an MCP
server
3. Create implementations for HTTP-SSE and StdIO servers, directly
copying the [MCP SDK
example](https://github.com/modelcontextprotocol/python-sdk/blob/main/examples/clients/simple-chatbot/mcp_simple_chatbot/main.py)
4. Add a util to transform MCP tools into Agent SDK tools
Note: I added optional caching support to the servers. That way, if you
happen to know a server's tools don't change, you can just cache them.
### Test Plan:
Checks pass. I added tests at the end of the stack.
---
#324#322#321
-> #320#319
### Summary:
1. Add the MCP dep for python 3.10, since it doesn't support 3.9 and below
2. Create MCPServer, which is the agents SDK representation of an MCP server
3. Create implementations for HTTP-SSE and StdIO servers, directly copying the [MCP SDK example](https://github.com/modelcontextprotocol/python-sdk/blob/main/examples/clients/simple-chatbot/mcp_simple_chatbot/main.py)
4. Add a util to transform MCP tools into Agent SDK tools
Note: I added optional caching support to the servers. That way, if you happen to know a server's tools don't change, you can just cache them.
### Test Plan:
Checks pass. I added tests at the end of the stack.
Currently, when we set `parallel_tool_calls=False` in the
`model_settings`, the responses API is called with `parallel_tool_calls
== NotGiven`, which defaults to true on the Response API side
(https://platform.openai.com/docs/api-reference/responses/create#responses-create-parallel_tool_calls).
This PR attempts to fix that.
```
agent = Agent(
...,
model_settings=ModelSettings(
parallel_tool_calls=False,
),
)
```
# Fix potential infinite tool call loop by resetting tool_choice after
tool execution
## Summary
This PR fixes an issue where setting `tool_choice` to "required" or a
specific function name could cause models to get stuck in an infinite
tool call loop.
When `tool_choice` is set to force tool usage, this setting persists
across model invocations. This PR automatically resets `tool_choice` to
"auto" after tool execution, allowing the model to decide whether to
make additional tool calls in subsequent turns.
Unlike using `tool_use_behavior="stop_on_first_tool"`, this approach
lets the model continue processing tool results while preventing forced
repeated tool calls.
## Test plan
- Added tests to verify tool_choice reset behavior for both agent and
run_config settings
- Added integration test to verify the solution prevents infinite loops
- All tests pass
## Checks
- [x] I've added new tests for the fix
- [x] I've updated the relevant documentation (added comment in code)
- [x] I've run `make lint` and `make format`
- [x] I've made sure tests pass
We don't really need mypy on 3.9 (unit tests would catch any real
issues), and it causes issues with the rest of this stack.
---
[//]: # (BEGIN SAPLING FOOTER)
* #324
* #322
* #321
* #320
* __->__ #319
### Summary:
This enables users to **use** MCP inside the SDK.
1. You add a list of MCP servers to `Agent`, via `mcp_server=[...]`
2. When an agent runs, we look up its MCP tools and add them to the list of tools.
3. When a tool call occurs, we call the relevant MCP server.
Notes:
1. There's some refactoring to make sure we send the full list of tools to the Runner/Model etc.
2. Right now, you could have a locally defined tool that conflicts with an MCP defined tool. I didn't add errors for that, will do in a followup.
### Test Plan:
See unit tests. Also has an end to end example next PR.
Eric Peter
Ali Waleed
Martin Jagusch
amri369
Steven Heidel
Rohan Mehta
pakrym-oai
Dmitry Pimenov
Dominik Kundel
Martín Bravo
apeccaud