# Simulations

> Watch real agents try to use your API, see where they drop off, and get the fix for each.

Source: https://docs.forgeintel.co/simulations

Rankings and Marketplaces show whether agents can find you. **Simulations** shows what happens next: real coding agents get a task your API can solve and try to complete it, from finding your service to paying for the call.

## Tasks

A task is a situation a user might hand their agent, such as "check tomorrow's weather in Lisbon", that one of your paid endpoints can solve.

1. Select **Draft tasks**. Forge writes up to five, covering every paid endpoint before it writes a second task for any of them. Each task's request is tried against your endpoint first, so no task is saved with values that never reach your `402`.
2. Review them. Each shows the endpoint it targets and how much it spells out: **All inputs given**, or **Some inputs left out** for the agent to work out. **Edit** or **Remove** any task.
3. Select **Run task** for one, or **Run all**. A task started while another runs is queued.

A simulation takes a few minutes. Forge emails you when the report is ready.

## How a run works

Each run is one agent in a fresh sandbox, with only the task's prompt to go on. Runs rotate through:

- **Agents:** Codex with GPT models and Claude Code with Claude models, at small and medium sizes.
- **Payment clients:** the AgentCash CLI, or a bare wallet with no payment tool, where the agent pays with libraries it installs or code it writes.

> No payment is ever settled. When an agent tries to pay, Forge records the request and the payment it would have made, and stops it.

## What you see

- **Agent success rate:** how many runs got to a valid paid request, by model and by payment client.
- **Where agents dropped off:** how many runs reached each step: reached the site, found the catalogue, chose the right endpoint, built a valid request, got the `402`, tried to pay.
- **Suggested fixes:** where failed runs stopped, why, and what to change on your side, ranked. Each cites the runs that prove it.
- **Coverage:** every paid endpoint, whether a task covers it, how often agents chose it, and whether it **checks inputs** before charging. An endpoint that returns a `402` for a request it can't fulfill **charges for broken calls**: agents pay and get nothing.

Open any run to see its full log: every step the agent took and why, the requests it sent, the request it was ready to pay for, and where it dropped off. **Simulation history** lists every finished run, newest first.