The parallel agent ticket-driven fix-and-land work loop

Issues are cleared by a fleet of background agents — one per ticket, each in its own isolated clone. The loop is: agree the fix, fork an agent, it repros-first and implements, then STOPS for a human to gate the commit; on approval it updates, re-tests, and fast-forward-posts. The method keeps many fixes in flight at once while every landed commit stays reviewed, tested, and ff-only.

One ticket per issue

Every issue is a numbered ticket e.g. todo/TOPIC/TOPIC-001.mkd, and the fix approach is agreed before any agent is forked.

One agent per ticket, in an isolated tree (worktree or clone)

Each agent gets its own tree WITH BEAGLE — either a store-backed worktree (cheap, shares the local .be) or a full independent clone — builds a debug tree there, and never touches another agent's files.

Implement, stop, report — the human gates the commit

An agent repros the bug, fixes it, gets the suite green, then STOPS without committing and reports for a decision.

Commit and ff-post, serialized

On approval the agent updates onto current trunk, re-tests, commits, and fast-forward-posts — one post at a time.

Reconcile against landed work

Because tickets land while others are in flight, each post rebases onto the current tip and merges any file two tickets both touched.

Keep the board current

todo/README.mkd lists all the active topics and what's going on there; todo/TOPIC/README.mkd lists active tickets (all bulletpoint 2 lines per entry). A fully-landed ticket gets removed from README.mkd, a closed topic from todo/README.mkd.

Hazards seen this session

A handful of traps recurred; each cost a stop-and-report, and each has a standing workaround.