Built-in loops
The standing automations every project ships with — the Organizer, Knowledge sync, and Skills.
Every project starts with standing automations pointed at one job: keeping the project's own memory true. You don't wire them — they're already there. What differs is how much autonomy each one gets before you say so.
| Loop | What it maintains | Default |
|---|---|---|
| Organizer | The Atlas — how every document is filed | On |
| Knowledge sync | The knowledge base entries themselves | Opt in |
| Skills | Your project's skills | Opt in |
The Organizer
It runs daily, and it's the one loop that's on from the start — because it only organizes and proposes, never authors. Each run rebuilds the Atlas: re-clusters the whole corpus into areas, keeps stable areas stable, retires empty ones. Then it curates each area:
- Identical entries fold automatically.
- Near-duplicates become merge proposals waiting in your Inbox.
- Contradictions get flagged rather than quietly resolved.
Its run history shows every decision: each area reviewed, each verdict, each model call. So a rebuild is auditable — when an entry moves, you can find out why.
Knowledge sync
Opt-in. It keeps the knowledge base current as the code changes: re-checks entries against the code they describe, proposes updates, and proposes new entries for the architecture nobody has written down yet.
A companion verify pass audits entries against reality, so stale claims get caught rather than trusted. That matters more than it sounds — a knowledge base every agent reads each turn is worse than useless when it's confidently wrong.
Skills
Opt-in, and it runs weekly. It curates your project's skills: prunes stale or duplicate ones, sharpens weak triggers, and adds genuinely reusable techniques.
Skills live in your repo, so this loop's audit trail is git history. Every change it makes is a diff you can read, review, and revert.
Why two of them wait for your yes
The split isn't arbitrary. A loop that only re-files and proposes can't damage anything: the worst case is a merge proposal you decline, so it can run unattended from day one. A loop that writes into your project's memory or into your repo changes what every future agent believes — so it waits until you turn it on.
That's also the honest way to earn trust. Turn on Knowledge sync, read a few runs, and decide from evidence whether you want it working while you sleep.
When these aren't the loops you need
These three are not special code. They're the same engine any standing loop uses — a role, a trigger, the collections it maintains, and a budget. Wire your own the same way: see building your own, and the event spine for what a loop can react to.