# Multi-profile comparison

This separate example evaluates three agent profiles on the same task:

- `filesystem-tools` exposes `read`, `write`, and `edit`;
- `bash-only` exposes only `bash`;
- the built-in `vanilla` agent profile exposes all registered Pi tools.

The deterministic integration test is the sanity check for arbitrary agent profile counts. It runs all three profiles without model requests and asserts all three pairwise comparisons:

```bash
npm run test:integration
```

Watch that exact test in the native Pi TUI:

```bash
npm run example:multi-scripted-live
```

Run the separate experiment with a real model:

```bash
npm run example:multi-live -- \
  --benchmark-preset hard \
  --model opencode-go/deepseek-v4-flash \
  --thinking high \
  --attempts 3
```

Use `--benchmark-preset easy`, `--benchmark-preset medium`, or `--benchmark-preset hard` to choose one of the three tasks. Use `--agent-profiles filesystem-tools,bash-only,vanilla` to change agent profile order or select a subset.