85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What are parallel tool calls?

Parallel tool calls are multiple tool calls a model returns in a single response, which the application can run concurrently and return together, so independent lookups finish in one round trip instead of several.

Checked

When a question needs the weather in two cities, a model that calls one tool per turn needs two round trips. A model that returns both calls at once needs one. Both Anthropic and OpenAI let models return several tool calls in a single response, and both let developers turn that off.

The model only proposes the calls. Whether they actually run at the same time is up to the application. Anthropic's documentation states that the API does not prescribe an execution order: calls can run concurrently, in sequence, or in a mix.

Running the calls

  • Concurrent when independent

    Independent, read-only operations are usually safe to run together for lower latency.

  • Sequential when they interact

    Tools with side effects, shared state or ordering requirements may be better run one after another.

  • Answer every call

    Anthropic requires one tool_result per tool_use, all together in the next user message and before any text. A call that was skipped still gets a result marked as an error with a short reason.

  • Match by ID

    In OpenAI's API each function call has a call_id, and each output is submitted with the matching ID.

  • Stricter for computer use

    Anthropic's computer use and browser use tools require batched actions to run in order and stop at the first failure.

Switching it off

In Anthropic's Messages API, disable_parallel_tool_use is set inside tool_choice, not at the top level. With tool_choice on auto it means at most one tool call per response; with any or a named tool it means exactly one. In OpenAI's API, setting parallel_tool_calls to false ensures zero or one tool call.

OpenAI documents two edge cases: on supported models from GPT-5 onward, built-in tools cannot be part of a parallel function-call batch, and one gpt-4.1-nano snapshot could return repeated calls to the same tool, for which OpenAI recommends disabling parallel calls.

Why parallel calls quietly stop

  • Results split across messages

    Anthropic calls this the most common cause: sending each tool result in its own user message teaches Claude to avoid parallel calls. All results belong in one message.

  • Weak prompting

    Claude 4 and later models make parallel calls by default when a request benefits from them, and a system prompt instruction can raise the rate further.

  • Dependent calls batched together

    If a call fails because its prerequisite had not finished, return the error and the model reissues it. Anthropic suggests telling the model to batch only calls that are independent of each other.

What it is confused with

Parallel tool calls happen inside one agent's turn. Agent parallelization runs several model calls or subagents side by side on parts of a task. Programmatic tool calling goes further: the model writes code that can run many tool calls concurrently, keeping their intermediate results out of the context window.

How Motiontic's engineers apply it

Motiontic's engineers mark each tool as safe or unsafe to run concurrently in the harness, run the safe ones together, and check traces for the formatting mistakes that make parallel calls fade over a long conversation.

Parallel tool call controls in the two main APIs

Parallel tool call controls in the two main APIs
Anthropic Messages APIOpenAI API
DefaultOnThe model may call several functions in one turn
Switch offdisable_parallel_tool_use inside tool_choiceparallel_tool_calls set to false
Effect when offAt most one call (auto) or exactly one (any or tool)Zero or one call
Linking resultstool_use_id on each tool_resultcall_id on each function_call_output
Result placementAll results in one user message, before any textOne output item per call

Common questions

Do parallel tool calls run at the same time?

Only if the application runs them that way. The model returns several calls in one response; the code decides whether to execute them concurrently or in sequence.

How are parallel tool calls disabled in Claude?

Set disable_parallel_tool_use to true inside the tool_choice object. With auto it limits Claude to at most one call per response.

Why did a model stop making parallel tool calls?

With Claude, the usual cause is returning each tool result in a separate user message. Returning all results together in one message keeps parallel calls working.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.