Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

Tool calls from one model step run one at a time, but the docs say they run in parallel

Open
  • Is something wrong?
#1,547 0 comments 1 reaction 1 assignee View on GitHub

Maintainers usually reply within 1 day

@tombeckenham is already working on this.

Since Sep 28, 2026.

Assessment

This issue has not been assessed yet.

Description

waiting-on: maintainer
TanStack AI version

@tanstack/ai 0.63.0. Also on 0.61.0.

Framework/Library version

Node 24, server-side chat(). No framework involved.

Describe the bug and the steps to reproduce it

When the model asks for several tools in one step, chat() runs them one after another. A step's tool time is the sum of its tools instead of the slowest one. With slow tools (a web search that takes 5 to 20 s), three calls in one step take 30 s instead of 10.

The docs say the opposite. Parallel Tool Execution (docs/tools/tool-architecture.md): three get_weather calls in one step, "All execute simultaneously, then LLM generates comparison."

What goes wrong

executeToolCalls (tool-calls.ts:929) loops for (const toolCall of toolCalls) and runs each server tool with yield* executeServerTool(...) (:1151, :1202), so each tool finishes before the next one starts. The engine hands it the whole step's batch (chat/index.ts:2161, :2349). ToolCallManager.executeTools (:350-396) has the same shape: await tool.execute(...) inside a for loop.

There's no option to change it and no comment giving a reason. The approval batch gate (:920) decides before anything runs, so it doesn't need the tools to run one at a time.

Repro

Sandbox linked below (output in the preview). It's the docs' own example: the model calls get_weather three times in one step, and each call takes 500 ms. The model is scripted, so no API key.

5ms start NYC
509ms start SF
1010ms start LA
1513ms done

Expected, per the docs: all three start at about 0 ms, and done at about 500 ms.

Fix

Start a step's server tools together and wait for all of them, keeping what the loop does today:

  • Results stay in call order, so the next model call sees them in the order it asked.
  • onBeforeToolCall still decides for each tool before it runs, and onAfterToolCall still runs for each result.
  • A tool's custom events still stream while it runs. They would interleave across tools, and each already carries its toolCallId.

Is running them one at a time intentional, for example for the middleware hooks or event order? If not, I'm happy to send a PR. If some apps rely on the current order, it could be an option instead of the default.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

https://codesandbox.io/s/259lf4

Screenshots or Videos (Optional)

No response

Do you intend to try to help solve this bug with your own PR?

No response

Terms & Code of Conduct
  • I agree to follow this project's Code of Conduct
  • I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.
Dominant language
TypeScript
Stars
3.1k
Forks
340
Avg merge
2d 10h
Merged PRs (30d)
175

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from TanStack/ai

All issues in TanStack/ai

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.