kaxil opened a new pull request, #73898:
URL: https://github.com/apache/airflow/pull/73898
Follow-up to #73587, which added Strands support through
`as_strands_tools()`. A tool that kept raising `ModelRetry`, or kept failing
argument validation, came back to the model as an error result every time, so a
Strands agent could loop without end. Raising from the tool does not help:
Strands catches any exception a function tool raises and hands it to the model
as another error. `as_strands_tools()` has not been released yet, so this
replaces it rather than deprecating it.
- **`AirflowTools` is now a Strands plugin.** Its `AfterToolCallEvent` hook
re-raises `ToolCallError`, which Strands raises out of the run as
`EventLoopException`, so a failure the model cannot fix fails the task and
Airflow's retry takes over.
- **Failures the model can correct are bounded by the tool's
`max_retries`**, counted as Pydantic AI counts them: once per tool per model
turn, so calls a framework runs concurrently in one turn count once.
- **Google ADK**: `AirflowTools` in `tools.adk` is an ADK toolset. A result
reaches the model as `{"result": ...}`, a failure it can correct as `{"error":
...}`, and any other failure raises `ToolCallError` out of the run.
- `airflow_tools()` moves to the shared toolset base, and `collect_tools()`
turns the toolsets and tools an adapter is given into one list.
- The LangChain bridge runs on the same interface: a correctable failure
becomes a `ToolMessage` with `status="error"`, and a tool that keeps failing
raises `ToolCallError`.
- `SandboxToolset` can own its sandbox from synchronous code with a `with`
block, which a task running a native agent needs. A tool call outside the block
is refused instead of provisioning a sandbox nothing would destroy. Attaching
to another task's sandbox works from both forms.
- `MCPToolset.airflow_tools()` raises: an MCP session belongs to the
Pydantic AI run that opens it, so a native agent should use its framework's own
MCP client.
**Why one neutral interface under both adapters?** Strands, ADK and the
LangChain bridge now share one conversion, one error model and one retry
budget, so an adapter for another framework is a thin mapping rather than a
copy of those rules.
**Why the Strands and ADK tests skip in CI.** Strands pins `mcp` below 2.2
and ADK pins OpenTelemetry at 1.42.1 or lower, both below the versions in
Airflow's development environment, so neither framework can be installed there.
Both integrations are marked experimental, and the docs list the versions they
are tested with (Strands 1.56.0, ADK 2.9.1).
An end-to-end Dag that runs every toolset through `AgentOperator`, a
Pydantic AI agent, Strands, ADK and LangChain passes with a real Claude model
and a Modal sandbox.
---
* Read the **[Pull Request
Guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#pull-request-guidelines)**
for more information. Note: commit author/co-author name and email in commits
become permanently public when merged.
* For fundamental code changes, an Airflow Improvement Proposal
([AIP](https://cwiki.apache.org/confluence/display/AIRFLOW/Airflow+Improvement+Proposals))
is needed.
* When adding dependency, check compliance with the [ASF 3rd Party License
Policy](https://www.apache.org/legal/resolved.html#category-x).
* For significant user-facing changes create newsfragment:
`{pr_number}.significant.rst`, in
[airflow-core/newsfragments](https://github.com/apache/airflow/tree/main/airflow-core/newsfragments).
You can add this file in a follow-up commit after the PR is created so you
know the PR number.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
