One agent per project
Each project has a written goal and one long-lived agent that owns it. The agent chooses its own next steps, hands the heavy work to subagents — coders, researchers, testers — and commits as it goes. Nobody hands it tickets.
An autonomous lab
Athanor
Every project gets an AI agent of its own. It wakes on a schedule, works within a fixed budget, checks itself with a critic and reports to one person in Telegram.
How it works
Woken on a schedule
A timer wakes each agent for a working session, mostly at night. Each wake resumes the same Claude session as the last, even after the budget guard stopped it. On top of it the journal — what is done, what comes next and why — is the durable memory, for when a session is cut, compacted or replaced.
A critic before “done”
Before a step counts as finished, a fresh critic reads the goal and the result and looks for what a demanding owner would reject. The agent fixes what it finds, or writes down why not.
A gate on the budget
Before every wake, a gate reads the subscription’s own usage meter. When the week’s share nears its cap, the wake does not start. The lab spends what it is allowed, and not a point more.
One conversation. The agents write to their director in a single Telegram chat: a result to look at, a question that blocks them, a short daily report. His replies land in the project’s inbox and are read at its next wake — where his word and the agent’s plan disagree, his word wins.
An athanor is the alchemist’s furnace: built to hold an even heat for as long as the work takes, feeding itself without being watched.