> ## Documentation Index
> Fetch the complete documentation index at: https://docs.extra-ai.co/llms.txt
> Use this file to discover all available pages before exploring further.

# Execution Limits

> Built-in guardrails that cap tool calls and iteration loops per run.

## Why limits exist

Without limits, a stuck model can loop forever — calling the same tool repeatedly, or an orchestrator delegating endlessly. Each loop costs money and time.

The engine enforces these caps at the runtime level, not through prompts. A prompt can ask the model to behave; only the runtime can guarantee it stops.

***

## Default limits

| Limit                        | Default | Meaning                                         |
| ---------------------------- | ------- | ----------------------------------------------- |
| `max_iterations`             | 20      | Model→tools→model rounds per node               |
| `max_tool_calls`             | 10      | Total tool calls per run (local + MCP combined) |
| `max_tool_calls_per_agent`   | 4       | Tool calls per agent per run                    |
| `max_child_agent_calls`      | 8       | Times an orchestrator can call child agents     |
| `allow_duplicate_tool_calls` | false   | Block identical repeated calls                  |

A **duplicate** is the same tool called with identical arguments in the same run.

These defaults apply automatically even with no `execution:` block in your YAML.

***

## Configuring limits

Add an `execution:` block to your spec. Any omitted key keeps its default.

```yaml theme={null}
execution:
  max_iterations: 30
  max_tool_calls: 20
  max_tool_calls_per_agent: 8
  max_child_agent_calls: 10
  allow_duplicate_tool_calls: false
```

Limits are validated at parse time — invalid values (non-integers, wrong types) fail `agentctl validate` with a clear error.

***

## What happens when a limit is hit

The run **degrades gracefully** — no exception, no crash.

* **Tool/child-agent limits and duplicates** — the call is blocked. The model receives a short message like *"Tool call limit reached, finish with what you have."* and naturally wraps up.
* **`max_iterations`** — the node's loop stops and returns the latest model response.

The engine logs a warning with the run id, limit name, and current count. It never logs prompts, tool arguments, or secrets.