AutoGen¶

Available · BaseTool per spec · pip install "agent-adapter[autogen]"

For AutoGen 0.7 (autogen-core, autogen-agentchat). AutoGen is in maintenance mode; Microsoft Agent Framework is its successor. For the ag2 package (AG2 1.x, a separate project), see AG2.

Functions¶

Function (`agent_adapter.integrations.autogen`)	Returns
`autogen_tool(spec, *, executor=None)`	one `autogen_core.tools.BaseTool` for `spec`
`autogen_tools(registry, *, executor=None)`	a tool for every registry tool whose `targets` include `autogen`
`AutoGenCall`	what `ctx.extras["autogen"]` holds: `call_id`, `cancellation_token`, `tool`
`INTEGRATION`	the integration object (`get_integration("autogen")`)

executor=None uses the default executor at call time (set_default_executor).

Use¶

executor = Executor(registry)  # hooks, approval, cache, ledger apply to every call
tools = autogen_tools(registry, executor=executor)
rate = tools[0]
print(rate.name, await rate.run_json({"region": "us"}, CancellationToken(), call_id="c1"))
# The workbench (what AssistantAgent uses) turns adapter errors into is_error results.
failed = await StaticWorkbench(tools).call_tool(
    "shipping_rate", {"region": "mars"}, call_id="c2"
)
print(f"is_error={failed.is_error}", failed.result[0].content)

In an agent:

agent = AssistantAgent("assistant", model_client=OfflineClient(), tools=tools)
result = await agent.run(task="Shipping to the US?")
print("answer:", result.messages[-1].to_text())

sequenceDiagram
  participant A as AssistantAgent
  participant W as StaticWorkbench
  participant T as BaseTool (adapter)
  participant E as Executor
  A->>W: call_tool(name, args, token, call_id)
  W->>T: run_json(args, token, call_id)
  T->>E: run_spec(spec, args, ctx(call_id))
  E-->>T: value, or raises
  T-->>W: value / exception
  W-->>A: ToolResult(content, is_error)

What AutoGen sees¶

`BaseTool` member	From
`name`, `description`	`spec.name`, `spec.description`
`args_type()`	`spec.args_model`
`schema["parameters"]`	the plain schema (refs inlined, titles dropped): `properties`, `required`, `additionalProperties` (default `False`); `strict: False`
`return_value_as_string(v)`	text as is, anything else as JSON

Behaviour¶

Topic	Rule
Validation	`run_json` sends the raw arguments to the executor; the tool's own args model validates them. `run(args_model_instance, token)` also works
Errors	Adapter errors raise from `run_json`; the workbench returns them as `ToolResult(is_error=True)` with the message (AutoGen's own error channel)
Call id	The workbench's `call_id` (the model's function-call id) becomes `ctx.call_id`
Approval	AutoGen has no per-tool gate: `confirm` tools ask the executor's `ApprovalPort`. Edited arguments from the port are validated again and applied. Use a port that blocks until it decides (`CallbackApproval`, a console or UI prompt): a `DeferredApproval` raises `ApprovalPending`, which the workbench (it catches `Exception`) reports to the model as a tool error instead of pausing the run
Cancellation	The `CancellationToken` is linked to the call: `token.cancel()` cancels the tool. `CancelledError` always propagates
Telemetry	AutoGen's tool span and `ToolCallEvent` are kept; the event shows no arguments for `confirm` tools
Components	Exported tools are not AutoGen components: `dump_component()` is not supported