# Limits

> The default limits on an agent turn, a scheduled firing, a function run, delegation, waits and recovery, and how to size a job against them.

These are the defaults. Whoever runs your Mindset deployment sets them, and you can't change them in the Studio, per org or per agent. What a run does when it reaches one is on [Runs](https://docs.mindset.ai/docs/runs).

## An agent turn

| Limit                                     | Default                | What it bounds                                                                                                    |
| ----------------------------------------- | ---------------------- | ----------------------------------------------------------------------------------------------------------------- |
| Time, when the caller waits for one reply | 90 seconds             | An API call that holds open until the answer is ready                                                             |
| Time, for streamed or delegated work      | 240 seconds            | A turn streamed back as it runs, or one another agent asked for                                                   |
| Model rounds                              | About 99               | How many times the agent can think and call tools. The underlying limit is 200 steps, and a round takes about two |
| Tokens                                    | 400,000                | What one turn can spend on the model                                                                              |
| Repeated calls                            | Stops at 5, warns at 3 | The same call with the same input and nothing new learned                                                         |

## A scheduled firing

- **15 min**: Time for the whole firing
- **12 turns**: Each up to 240 seconds
- **1.2M tokens**: For the whole firing

## A function run

| Limit                                  | Default                         | What it bounds                                                                                 |
| -------------------------------------- | ------------------------------- | ---------------------------------------------------------------------------------------------- |
| Time, when an agent or script calls it | 5 minutes                       | The whole run. Retries don't extend it                                                         |
| Calls to outside systems               | 1,000                           | Every operation or model call. Each item in a list counts once, however many attempts it takes |
| Items in a list a step repeats over    | 1,000                           | A longer list is refused before any call goes out                                              |
| Items in flight at once                | 8                               | How many of those calls run at the same time                                                   |
| Attempts per step                      | 1 unless you set more, up to 10 | Retries on the failures you choose, such as rate limits and timeouts                           |

## Delegation and MCP

| Limit                                                        | Default                                                      |
| ------------------------------------------------------------ | ------------------------------------------------------------ |
| Calls from one run to other agents                           | 8, shared across the whole chain                             |
| Wait before a delegated call becomes a task to check back on | 30 seconds. The work keeps going to its own 240 second limit |
| A call to a tool on an MCP server                            | 120 seconds                                                  |

## Waits and recovery

| Limit                                                 | Default                                                      |
| ----------------------------------------------------- | ------------------------------------------------------------ |
| Shortest wait a script phase can declare              | 5 minutes. A shorter one is refused when the script is saved |
| A parked run is closed as abandoned after             | 31 days                                                      |
| A run whose server stopped is picked up again after   | About 5 minutes                                              |
| A run that keeps failing is closed as abandoned after | An hour, or 10 failures in a row                             |

## Size a job against them

Count what one item costs. The invoice exceptions agent works through four phases per invoice:

| Phase   | Rough cost                     |
| ------- | ------------------------------ |
| Gather  | 2 operation calls              |
| Compare | 1 function run                 |
| Assess  | 2 or 3 knowledge base searches |
| Prepare | 1 write operation              |

One invoice fits easily in one turn. Four thousand in one run don't: time runs out first, and a function step stops at 1,000 items. Change the unit of work instead. For small work per item, repeat a function step across a list of a few hundred. For bigger work per item, run the agent once per invoice, started by your finance system [through a webhook](https://docs.mindset.ai/docs/start-it-from-your-server), so each run has its own record and can be re-run alone. Size a batch for a slow morning: if it fits only when every system answers at once, it doesn't fit.