# OS Manager caps tool calls per run and per specialist

> Agno's OSManager sets tool_call_limit=8 for each coordinator run and member_tool_call_limit=24 for each specialist invocation. Pass None to either one for unlimited tool use.

- Published: 2026-10-20
- Author: Ashpreet Bedi
- Categories: Changelog
- Canonical: https://www.agno.com/blog/os-manager-caps-tool-calls-per-run-and-per-specialist
- Markdown: https://www.agno.com/blog/os-manager-caps-tool-calls-per-run-and-per-specialist.md

Agno's `OSManager` passes its two new settings to the tool call limit that agents and teams already have, so the lead team counts every ordinary tool execution in a run, delegations included, and each specialist counts its own from zero on every call. OS Manager is the AgentOS team that checks, runs and builds your components.

A simple question could grow into repeated docs searches, delegations and source reads. Now a call past the limit is skipped and returned to the model as a refusal, and a shared prompt tells every manager role to stop there and report verified progress and gaps.

We built the default manager on both releases and printed each `tool_call_limit`:

```text
# real output, default OSManager()

# Before
OS Manager: tool_call_limit=None
  Platform Builder: tool_call_limit=None
  Platform Manager: tool_call_limit=None

# Now
OS Manager: tool_call_limit=8
  Platform Builder: tool_call_limit=24
  Platform Manager: tool_call_limit=24
```

Then we gave the lead team a scripted stand-in model that asks for the same lookup 12 times. Earlier releases ran all 12. This release ran 8 and sent the model this message:

```text
# real output, scripted stand-in model
Tool call limit reached. Tool call focused_lookup not executed. Don't try to execute it again.
lookups executed: 8 of 12 requested
```

The prompts got tighter too. The lead and Platform Manager read an ambiguous term such as "learning machines" as an Agno concept first, cite the docs they used and keep a basic explanation within 200 words. Every role keeps framework behavior, local configuration and verified execution apart.

Raise either limit for larger builds, or pass `None` to lift it. Application agents keep their own limits.

```python
OSManager(model=model, tool_call_limit=16, member_tool_call_limit=48)
```

See the [cookbook](https://github.com/agno-agi/agno/blob/main/cookbook/05_agent_os/29_manager/README.md), and learn more about [tool call limits](https://docs.agno.com/tools/tool-call-limit) in the documentation.