Your Claude or Codex subscription is the engine. TeamX is the control plane. Flip to autonomous mode and your subscription becomes a fully autonomous dev team — tasks that spec themselves, ship in safe lanes, and document themselves. Auto Docs, Blast Radius, decisions, test results, budget. All in one loop. Flip a switch and the agent takes the safe wheel.
Default. Agent does nothing on its own. Every task waits for you to pick it up. Same behavior as the kill switch.
Agent surveys the product on a cron and drafts proposals into an approval queue. Nothing ships without you clicking approve. You get a daily digest of what it found.
Research, plus safe-lane work auto-ships under guardrails. Risky work (anything touching forbidden surfaces, public pages, or flagged by Blast Radius) still waits for approval.
Hard cap on tasks shipped per day (default 5) and max files touched per task (default 10). Over budget = parked, never skipped silently.
Pricing, billing, features_schema, auth — untouchable by default. Any proposal hitting these auto-parks with a reason.
Every ship dispatch runs through compute_blast_radius. Public surface or recent breaking change = parked unless the proposal is explicitly safe_lane.
Paste your Claude /usage screen once a day. When your weekly quota drops below the floor (default 20%), ship loop pauses until next reset.
Each capability is a research + execute pair. Toggle them per team. Shipping a new capability is just “add a row + two endpoints”.
Regenerates stale docs and backfills missing doc→epic links. The first capability we shipped because it’s the safest.
Surveys rising feedback votes, repeated manual work, competitor signals → drafts new epic proposals with business case.
Watches Sentry, failed workflows, failing tests → safe-lane fixes auto-ship, risky ones draft for approval.
Lighthouse + a11y + dead-link checks → fixes missing alt text, broken links, empty states.
Picks model per task size/risk/time-of-day/budget. Sonnet in peak hours, Opus for heavy lifts, extractive fallback when quota is tight.
Spots repeated patterns in activity + messages → proposes automations that would save time.
Settings → Autonomous Mode. Toggle to research. The agent starts surveying immediately on the 30-minute cron.
Each capability has an on/off toggle. Documentation Sweeper is enabled globally; others you opt into per-team as they ship.
Tasks per day, max files per task, forbidden surfaces, weekly quota floor. Conservative defaults ship with every team.
Run /usage inside Claude Code, copy the dialog, paste into the Settings panel. Autonomous Mode uses this to stop shipping when your weekly quota is low.
Every morning you get one message in #autonomous with counts and samples of what ran overnight. Researched, shipped, proposed, parked, failed. Two-minute read.
Once you trust the digest for a few days, flip to ship. Safe-lane work auto-ships. Everything risky still waits for you.
One button in Settings. Flips mode to off and stamps autonomous_halt_at. In-flight runs check this before every step and exit cleanly. Typically stopped within seconds.
Every morning at 14:00 UTC, one message lands in the #autonomous channel. Researched, shipped, proposed, parked, failed. Samples of each. Two-minute read over coffee.
Each capability labels its proposals safe_lane=true when the action is reversible, low blast radius, and not touching forbidden surfaces. Documentation Sweeper marks doc regeneration and doc-link insertion as safe-lane because both are idempotent and easy to undo. Capabilities decide what counts; the guardrail layer still blocks if the proposal’s blast radius lights up.
Every decision is in autonomous_run_log with capability, step, outcome, reason. Every shipped action is in autonomous_proposals with execute_result. Rolling back = it depends on the action. For doc regeneration, hit Regenerate again. For doc links, delete the work_item_docs row. Bigger capabilities will ship with capability-specific rollback helpers.
Yes, and the weekly quota floor exists to prevent that. Paste /usage once a day. When weekly_pct_used crosses 100 minus your floor (default 20%), ship loop pauses. Research loop is cheap — reads from the DB, no LLM calls — so it keeps going.
A cron runs without understanding what’s safe. Autonomous Mode reads the full TeamX graph — Blast Radius for impact, Institutional Memory for decisions, Auto Testing for verification, Budget for cost — before doing anything. The capabilities only propose; the core loop decides whether to ship.
Big red Kill switch button in the Settings panel, or flip mode to off. Both set autonomous_halt_at. In-flight runs check this before every step and exit cleanly at the next safe point (typically within seconds).
In research mode, yes — that’s the whole point. Product Research capability surveys your signals and drafts new epic proposals with cited sources. You approve or reject each one from the digest. In ship mode these still stay in the proposed queue; only safe-lane proposals auto-ship.
Settings → Autonomous Mode → Research. No code to write. You’ll get your first digest in about 24 hours.