Launched this week
The Autonomous Stack
Production-tested architecture for autonomous Claude agents
30 followers
Production-tested architecture for autonomous Claude agents
30 followers
A production-tested reference architecture for autonomous Claude agents. 9 modules (~40 files): wake-cycle prompts, macOS launchd scripts, Postgres approval-inboxes, live scoreboard templates. Built from 6 months of draining bankrolls in public.
Six months of Claude agents draining real bankrolls in public is rare production data for this space. Curious about the state boundary at wake-cycle transitions: if a running task spans a wake boundary, does it resume from persisted state or restart cold, and have you seen cases where the persisted state itself got corrupted enough that the agent made worse decisions on resume than on a fresh start?
@myultidev Each wake is a total cold start. launchd kills and relaunches the process, so nothing survives in memory and tasks don’t span wake boundaries.
The architecture is split:
Queue file: Pending tasks.
Postgres: Completed history.
On every wake, the agent pulls its journal/action history from the DB to pick the top queue item fresh.
The Risk: Corruption is real. I’ve seen an agent log a successful outreach that actually bounced; the next three wakes inherited that false state and skipped the lead. It took two days to catch.
The Fix: Every "Class-A" action (external sends or state mutations) now requires an artifact logged to Postgres with a verifiable_at timestamp. It doesn't make corruption impossible, but it narrows the window from "silently wrong for weeks" to "visible within a few cycles."