Add a per-run limit for tool calls #1202

Activity

1. Lokesh118 commented on Oct 24, 2026

   

   Lokesh Nelapati (Lokesh118)

   on Oct 24, 2026

   More actions

   Hi team!

   I'd like to take this up if nobody is currently working on it.

   Proposed Approach:

   1. Config Option: Add a new pointer field in agent/harness/toolautocall/autocall.go under Config:

      // MaximumToolCallsPerRequest (or MaximumFunctionCallsPerRequest) is the maximum
      // number of cumulative tool invocations allowed in a single run.
      // When nil (the default), no call-count limit is enforced.
      MaximumToolCallsPerRequest *int

      (Since providers like openaiprovider, anthropicprovider, etc. accept config.ToolAutoCall *toolautocall.Config, this immediately becomes configurable across all adapters).

   2. Accounting: Track cumulative tool invocations across turns within autocall.run().

   3. Budget Exhaustion Behavior: As suggested in the issue, allow an in-progress batch from the current iteration to finish. Between iterations, if cumulative invocations meet or exceed the limit, strip schema tools for the next round via prepareOptionsForLastIteration(opts) so the model gracefully wraps up with a final text response.

   4. Validation & Unit Tests: In autocall_test.go:

      • Validate negative values (returns configuration error).
      • Behavior when set to 0 vs nil (unlimited).
      • Cap triggered within a multi-call turn (allowing current batch to complete, then wrapping up).
      • Exact limit match vs overshoot across multiple iterations.

   Do you have any preference between MaximumToolCallsPerRequest vs MaximumFunctionCallsPerRequest? If this approach looks good to you, I'll go ahead and submit a PR!
2. Yut0Miura commented on Oct 26, 2026

   

   Yuto (Yut0Miura)

   on Oct 26, 2026

   Author

   More actions

   Sounds good. I'd go with MaximumFunctionCallsPerRequest to stay consistent with Python's max_function_calls.

   A few things worth covering:

   • nil means unlimited, 0 means no automatic function calls, but the model can still return a final response.
   • Count actual execution attempts, including failures, not just tool calls requested by the model. Provider-handled calls shouldn't count.
   • Include approved tool calls in the count, since they use a separate execution path.

   Letting the current batch finish even if it goes over the limit is fine, as long as no new batch starts afterward.

   The rest of the approach looks good. Go ahead with the PR.
3. Lokesh Nelapati (Lokesh118)

   added 8 commits that reference this issue on Oct 29, 2026

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   29f7445

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   f6f93ad

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   28297a7

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   4d2bb3d

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   571f436

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   7393a6d

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   20faef3

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   c08068c
4. Lokesh Nelapati (Lokesh118)

   added a commit that references this issue on Nov 14, 2026

   Add MaximumFunctionCallsPerRequest to toolautocall (microsoft#1202)

   13abe7b