Jerry

The problem

Running many agent tasks at once usually means runaway cost and no visibility. We wanted the leverage of parallel agents with the cost profile of a careful manual session — and a board we could actually watch.

What we built

1. 01

   One subprocess per task

   Each task gets its own Claude subprocess for its full duration, in an isolated git worktree, with its cache kept warm. tmux owns process state; the driver is pure plumbing at zero token cost.

2. 02

   Cost derived from the transcript

   Status, cost, and diff are read straight from Claude Code's session JSONL — deduped by message id so streamed duplicates don't triple-count. A cost cap kills a task that runs away.

3. 03

   A board you can watch

   A kanban shows every task as running, needs-input, or done, with live cost and a diff chip. Push notifications fire on the transitions that matter.

How we measured it

Throughput versus a single manual session, median cost per task, and cache hit rate — the same numbers the dashboard surfaces in real time.