raw/fleet-of-agents-sermakarevich.md

Raw capture — raw/fleet-of-agents-sermakarevich.md

url: https://github.com/sermakarevich/claude/tree/main/usage_pattern/fleet_of_agents date_fetched: 2026-07-05 backfilled: true

A tutorial that builds up to running a fleet of Claude coding agents in parallel — one capability per step. Stop at any step and you have a working setup. Start with a single agent that writes its work to disk; end with multiple parallel agents consuming a shared queue and asking you questions when stuck.

Step	Adds	When to stop here
Step 1 — materialize output	Agent writes task context ( `TASK.md`/`PLAN_AND_PROGRESS.md`/`FINDINGS.md`) to`.claude/tasks/<task_id>/`so work survives across sessions.	One agent, manual driving, but every task is resumable.
Step 2 — TODO/DONE queue	On-disk queue: append to `TODO.md`, agent picks the top entry and logs to`DONE.md`.	You manage a backlog but still drive each iteration.
Step 3 — loop.sh	`./loop.sh`invokes`claude -p`repeatedly until the queue is empty.	One unattended agent chewing through your TODO.
Step 4 — beads	Replaces `TODO.md`/`DONE.md`with`bd`; the loop atomically claims tasks; multiple`./loop.sh`instances can run in parallel.	A fleet of parallel agents on a shared queue.
Step 5 — Q&A blocking	Agent can ask you a question (writes to `Q&A.md`+ marks the task blocked); the loop skips blocked tasks until you answer.	A fleet that can stop and ask instead of guessing.