# Tasks, claims and checks

An agent that says "tests pass" is making a claim. BrigadeStack can turn the claim into a record: it runs the real check itself and writes down the exit code. Checking is optional. Shared memory works without it, and [Modes](/docs/concepts/modes) shows how to turn it off.

## The three parts

**Tasks.** A task is a piece of intent with acceptance criteria. File one with `brigadestack work task add`, list them with `brigadestack work tasks`, and start a session from the brief that `brigadestack work brief` prints.

**Claims.** When several agents work on one repo, a claim says who has a task. `brigadestack work claim --actor <name>` takes one, `--next` takes the highest-priority ready task, and `brigadestack work release` frees it. Two agents cannot hold the same task.

**Checks.** `brigadestack work verify run` runs the real command and records its exit code, the git state and the code-graph delta in a plain file, `receipt.json`. The model never grades its own work. The record then lands in the Evidence Ledger, so the next brief starts from what actually happened.

```bash
brigadestack work task add "fix the login redirect" --acceptance "tests pass"
brigadestack work claim --actor codex --next
brigadestack work verify run --command "npm test"
```

## What happens to the record

Reviewed notes can become durable memory, but an agent writes one only when the session learned something the repo, git history or a card lacks. Otherwise it ends with `Nothing durable: <reason>`. `brigadestack outcome` scores the skills and cards that produced a result, but only from checks run with `--manifest`, a registered verifier manifest. A check run with `--command` is an audit record. It is kept in the ledger and never scores. [Score skills and agents](/docs/guides/score-skills-and-agents) shows how.

## Where each part lives

| Part | Command | What it leaves behind |
|---|---|---|
| Task | `brigadestack work task add`, `brigadestack work brief` | task ledger |
| Claim | `brigadestack work claim`, `brigadestack work release` | holder on the task |
| Check | `brigadestack work verify run` | `receipt.json` with the exit code (audit-only for `--command`) |
| History | `brigadestack evidence search`, `brigadestack outcome rank` | Evidence Ledger, `memory/outcome/` |

BrigadeStack does not host your agents. The commands above write plain files when called, and hooks in a wired repo send checks through them. The only background work is the jobs `care install` schedules, such as the ingest sweep and a daily memory check. To hand a task to another vendor's agent, see [Run a task on another agent](/docs/concepts/bounded-runs).