[Zakaria Khchiche (Zakariakhchiche)]

Motivation & Context

allow_multiple_tool_calls=False is only translated to Anthropic's tool_choice.disable_parallel_tool_use when tool_choice is set. With tools but no tool_choice, the setting is silently dropped. The function-invocation loop also resets tool_choice="required" to None after the first iteration, so every later model call in the loop re-enables parallel tool use.

Description & Review Guide

• What are the major changes?
  • In _prepare_tools_for_anthropic, when tool_choice is unset, allow_multiple_tool_calls is set and tools are present, send {"type": "auto", "disable_parallel_tool_use": not allow_multiple_tool_calls}. auto is Anthropic's default mode when tools are provided. The function-invocation loop is not touched.
  • Regression tests: a unit test on _prepare_options, and a function-loop test (tool_choice="required" + allow_multiple_tool_calls=False, three mocked BetaMessage responses) that checks disable_parallel_tool_use on every request.
• What is the impact of these changes? The parallel-tool setting now applies to every request, including later function-loop iterations. Requests without tools or without allow_multiple_tool_calls are unchanged.
• What do you want reviewers to focus on? Gating on tools only (not mcp_servers), to avoid sending a tool_choice on requests without client tools.

Related Issue

Fixes #8979

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

Review effort: Lite
Findings: 2

Open (2)

• This assertion is overly strict (full dict equality), making the test brittle if tool_choice… · New
• These assertions also rely on exact dict equality for tool_choice, which can cause unnecessary… · New

What changed in this PR

This PR ensures allow_multiple_tool_calls is honored for Anthropic requests even when tool_choice is unset, by explicitly sending tool_choice={"type":"auto", "disable_parallel_tool_use": ...} when tools are present. It also adds regression tests to confirm the setting persists across function-invocation loop iterations where tool_choice="required" may be reset internally.

Changes:

• Update _prepare_tools_for_anthropic to propagate disable_parallel_tool_use via default "auto" mode when tool_choice is None and tools are present.
• Add unit coverage for _prepare_options to ensure allow_multiple_tool_calls=False disables parallel tool use without an explicit tool_choice.
• Add a function-loop regression test asserting disable_parallel_tool_use is sent on every loop request.

File	Description
python/​packages/​anthropic/​agent_framework_anthropic/​_chat_client.py	Propagates allow_multiple_tool_calls into Anthropic tool_choice even when user didn’t set tool_choice, but tools are present.
python/​packages/​anthropic/​tests/​test_anthropic_client.py	Adds regression tests for tool_choice/parallel-tool behavior in both option preparation and multi-iteration function loops.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[Zakaria Khchiche (Zakariakhchiche)]

Notes on the automated review: the new tests now assert the tool_choice fields individually, like the neighbouring _prepare_options tests (pushed). They still fail on main and pass with the fix.

[Eduard van Valkenburg (eavanvalkenburg)]

Zakaria Khchiche (@Zakariakhchiche) there are still unresolved comments, could you have a look?

[Zakaria Khchiche (Zakariakhchiche)]

Thanks Eduard van Valkenburg (@eavanvalkenburg)! Both open threads were automated review suggestions that were already applied: the tests assert the tool_choice fields individually (lines 1434 and 1487-1493). I've replied on each and resolved them. Happy to change anything else.