I built DevClaw, an OpenClaw plugin that turns each Telegram group into an isolated, autonomous dev team. Orchestrator, DEV, and QA all running on their own. I use it for all my development now. Here’s what it does:
∙ Autonomous dev/QA loops: DEV writes code, QA reviews it, failures loop back to DEV automatically, no human in the loop
∙ Multi-project isolation: each Telegram group is a separate project with its own queue, workers, sessions, and state
∙ Smart model selection: picks the cheapest model per task (Haiku for typos, Sonnet for features, Opus for architecture)
∙ Session reuse: workers keep their sessions across tasks instead of re-reading the entire codebase every time (~50K tokens saved per pickup)
∙ Token-free scheduling, the heartbeat that scans queues and dispatches work runs on pure CLI calls, zero LLM tokens
∙ Atomic operations, label transition, state update, session dispatch, and audit log in one call with rollback on failure
∙ Issues as source of truth, everything runs on GitLab/GitHub issues, no internal database
These compound to roughly 70% token savings versus running one big model with fresh context each time.
Keen to hear from anyone running autonomous coding agents, especially what you’ve found works for model selection and session management.
I built DevClaw, an OpenClaw plugin that turns each Telegram group into an isolated, autonomous dev team. Orchestrator, DEV, and QA all running on their own. I use it for all my development now. Here’s what it does:
∙ Autonomous dev/QA loops: DEV writes code, QA reviews it, failures loop back to DEV automatically, no human in the loop
∙ Multi-project isolation: each Telegram group is a separate project with its own queue, workers, sessions, and state
∙ Smart model selection: picks the cheapest model per task (Haiku for typos, Sonnet for features, Opus for architecture)
∙ Session reuse: workers keep their sessions across tasks instead of re-reading the entire codebase every time (~50K tokens saved per pickup)
∙ Token-free scheduling, the heartbeat that scans queues and dispatches work runs on pure CLI calls, zero LLM tokens
∙ Atomic operations, label transition, state update, session dispatch, and audit log in one call with rollback on failure
∙ Issues as source of truth, everything runs on GitLab/GitHub issues, no internal database These compound to roughly 70% token savings versus running one big model with fresh context each time.
Keen to hear from anyone running autonomous coding agents, especially what you’ve found works for model selection and session management.
Code’s all here: https://github.com/laurentenhoor/devclaw