HOW IT WORKS

The runtime, without the hand-waving.

An agent is a loop around a model with tools attached, permissions in front of it, and a platform underneath that keeps the loop alive across days, restarts and schedules. Here is each part of that, in order.

01
The turn loop

Every turn assembles a prompt from the agent's soul, its instructions, its available tools and the session history, sends it to the model you configured, and executes whatever it calls. Tool results land back in the transcript and the loop continues until nothing is left to call.

1 · prompt assembledsoul + instructions + tools + history
2 · model respondstext, or a tool call
3 · tool runsresult appended to the transcript
4 · loopuntil the agent has nothing left to call
02
Capabilities, then permissions

An agent has roughly forty built-in tools: shell over SSH, background commands, file read/write/edit, its own private workspace plus a shared account workspace, delegation, reminders, a shared todo list, and outbound messaging. Nothing is implicit — hosts must be verified per account and allow-listed per agent.

kestrel — permissions
ssh hosts edge-1, edge-2, edge-3 verified
may call bosun, lookout
skills certs, backups, host-health
denied billing-*, prod-db
03
Delegation

An agent can spawn a subagent scoped to one piece of work, message a peer specialist directly, and report a result back to whoever called it. The transcript interleaves all of it: your messages, tool calls with arguments and results, peer messages, subagent reports and system events.

quartermasteragent-to-agent threadRunning
youGet the edge fleet ready for the audit on Friday.
agents.messagekestrel — "cert + patch status, all edge hosts"1.2s ✓
agents.messagelookout — "summarise the audit checklist"0.9s ✓
kestrelquartermaster · report
All three hosts patched. edge-3 cert rotated today; chain verified by bosun.
lookoutquartermaster · report
Checklist has 14 items. Four need evidence we do not currently collect.
quartermasterFleet is audit-ready except four evidence items. I have assigned those to navigator as a standing mission so they are collected daily until Friday.
04
Durability

A mission is checkpointed. If the runtime restarts mid-tool, the session resumes rather than dies — the retry is ours to handle, not something you wake up to.

09:00:04mission weekly-sweep fired
09:02:11subagent bosun working
09:02:12runtime restarted
09:02:19session resumed from last checkpoint
09:04:47mission completed
05
Schedules and wake-ups

Missions fire on cron, on inbound webhooks, on a channel message, or on a heartbeat the agent uses to check in with itself. Plugins can intercept a fire to annotate it — or cancel it outright.

standing missions
0 6 * * 1weekly host health sweep
*/30 * * * *triage inbound support mail
heartbeat 4hself check-in — anything waiting on me?
webhooksentry alert → wake kestrel
Bring your own models
Register a provider with your key, add its models with per-token pricing, and point each agent at whichever one suits its job.
Metering built in
Spend is recorded per session, per agent and per model, with quotas and history — so a runaway mission is visible, not a surprise.
Reachable from outside
Route a Telegram or WhatsApp channel to a specific agent and it answers there with the same tools and permissions.

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.