Autonomous agent platform
A team of AI workers you brief from Telegram — and get finished work back, not a draft.
- Assign by voice or text from your phone: triage, execution, quality review, report with cost
- Recovers stuck tasks on its own and reroutes to a backup engine when the primary hits its limit
- Isolated project contours: every team has its own memory and its own charter
The problem
A single assistant is bottlenecked on the human: it waits to be explained to, checked and reminded. I needed a loop that finishes the job itself — including reviewing its own quality and honestly reporting failure when it fails.
How it works
Twelve services in Docker Compose behind Caddy: a task queue, an orchestrator that triages by difficulty, specialists (coder, analyst, designer), a separate reviewer agent, and a bridge to Claude Code. Telegram is not a chat here but a control panel: a forum topic maps to an agent or a project team, and the answer lands back in the same thread.
The hard part
Autonomy cost far more than intelligence. A watchdog re-opens stalled tasks through an atomic claim, distinguishes a transient error (retry) from an exhausted quota (park, don't retry), and on a subscription limit transparently reroutes the turn to a backup engine. A separate battle was cost: 73% of token spend turned out to come from resumed sessions, and the fix was not a better model but a different session shape.
Evidence
- 53/53 unit tests; verification through end-to-end runs on real tasks, not mocks
- The review loop proven live: rework → re-verification → an honest 'not approved' status
- Limit failover proven by forcing it: the turn moved to the backup engine and the task completed
- Per-task cost is measured and printed on the result card