USE CASES

Three missions,
end to end.

Not capabilities lists. Three real shapes of work, with the transcript that produced them — including the parts where an agent handed the job to somebody better suited.

MISSION 01

Cert rotation across an edge fleet

crew: quartermaster, kestrel, bosun, navigator

A standing mission fires every Monday at 06:00. Nobody is watching. The infra operator sweeps host health, discovers a cert that is about to expire, rotates it, and then refuses to certify its own work — it spawns a subagent to verify the chain independently, because its instructions say so.

A cert rotated, verified by a second agent, and the next rotation scheduled — with nobody awake for any of it.
kestrelscheduled run · Mon 06:00 · $1.14Running
◷ systemscheduled prompt fired — "weekly host health sweep"
▸ shell.runhost=edge-1 · df -h; uptime; certbot certificates0.8s ✓
▸ shell.runhost=edge-2 · df -h; uptime; certbot certificates0.9s ✓
kestreledge-3's cert expires in 6 days. Rotating now; dry run first per instructions.
▸ shell.runhost=edge-3 · certbot renew --dry-run2.1s ✓
▸ shell.runhost=edge-3 · certbot renew && systemctl reload nginx4.7s ✓
▸ agents.spawnbosun — "verify chain + OCSP on edge-1..3, do not trust my word"✓
bosun → kestrel · subagent report
Chain valid on all three. edge-3 was still serving the old intermediate from cache until the reload; verified after. OCSP stapling responds on all hosts.
▸ reminders.set2026-11-02 09:00 — begin next rotation, edge fleet✓
kestrelSweep done, edge-3 rotated and independently verified. edge-2 is at 88% disk — added a todo and a reminder to prune Thursday if it passes 92%.
MISSION 02

A research brief with citations

crew: quartermaster, lookout, two subagents

A founder asks a question that would take an afternoon. The orchestrator splits it, runs two readers in parallel on different sources, and reconciles what they disagree about instead of averaging it away.

Eighteen cited sources and one flagged contradiction, in six minutes, for two dollars and change.
quartermasterhuman chat · 6 min · $2.31Running
youWhere has the state of the art on agent memory compaction actually landed? I want sources, not vibes.
quartermasterSplitting this: one reader on papers and repos, one on practitioner writeups. I will reconcile.
▸ agents.spawnlookout — "papers + repos, last 9 months"✓
▸ agents.spawnlookout-2 — "practitioner posts, changelogs, issue threads"✓
lookout → quartermaster
11 sources. Consensus is moving from single-pass summarisation to structured, iteratively refined state. Two papers measure retention; both favour explicit goal/progress framing.
lookout-2 → quartermaster
7 sources. Practitioners report the same, plus one caveat the papers miss: refinement quality degrades if the summary is rewritten from scratch each time rather than amended.
quartermasterWritten to workspace: brief.md, 18 sources, with the disagreement flagged rather than smoothed. Short version — structured state beats prose summaries, and amending beats rewriting. Pi Compactor in the marketplace does both, which is a reasonable thing to try before building anything.
▸ files.writeworkspace/research/brief.md · 4.2 KB✓
MISSION 03

Inbound triage that never sleeps

crew: navigator, triage, kestrel

A support address and a Sentry webhook both route into the crew. Every thirty minutes the triage agent sorts what came in; anything that smells like infrastructure gets handed to the operator, and anything that needs a human gets a message on Telegram rather than a dashboard nobody opens.

An incident found, fixed and communicated in under an hour, with exactly one thing escalated to a human.
triagecronjob · */30 * * * * · $0.38Running
◷ systemwebhook — sentry — TLS handshake failures, checkout-3, 41 events
▸ mail.listsince=14:00 · 9 new0.4s ✓
triageSix are routine. Two are the same TLS report from customers. One is a billing question — out of my scope.
▸ agents.messagekestrel — "TLS handshake failures on checkout-3, customer-visible"1.1s ✓
kestrel → triage
checkout-3 had yesterday's bundle. Rotated and reloaded; errors stopped at 14:41. Root cause was a failed deploy hook, not the cert.
▸ telegram.send#ops — "checkout TLS resolved 14:41. Billing question from Acme needs a human."0.5s ✓
triageReplied to both customer threads with the resolution time. Billing question left for a person and flagged in the shared todo.
The shape is always the same
Something wakes an agent — a schedule, a webhook, a message, a heartbeat. It works, delegates what it should not do alone, and reports.
Delegation is where quality comes from
The agent that made a change is not the agent that verifies it. Two readers on two sources beat one reader on both.
Escalation is a feature
A crew that knows what to hand to a human is worth more than one that guesses confidently.

Crews of expert agents that finish the work.

Create an account, add your model key, and write your first agent's soul. The rest of the crew can arrive in one click.

Assemble your crewRead the docs