What it is
The BTF4E Agent Toolkit is the set of scripts and notes we use to run several AI coding agents at once. Each agent gets its own name and identity, posts what it's doing to a shared message channel, and works on its own branch. Helper scripts show a dashboard of who is doing what, predict merge conflicts before they happen, and launch or coordinate agents one after another or in parallel. A roadmap file lists the work agents can pick up. It also includes Flywheel, an earlier experiment in which scheduled agents read a folder of project notes each day, wrote a morning plan, pushed a few tasks forward and checked in every few hours. Flywheel ran from March to September 2026 and has since been retired in favor of a simpler ticket system.
How it works
- Give each agent an identity and start the message broker.
- Launch agents, one at a time or several in parallel.
- Agents announce their work; you watch the status dashboard and message log.
- Before merging, run the conflict predictor to spot overlapping changes.
- Clean up stale branches when a phase is done.
What exists today
- Git helpers: agent status dashboard, identity setup, conflict prediction, repository survey, branch cleanup
- Messaging scripts: start the broker with saved history, announce, view messages
- Agent scripts: launch, serial and parallel orchestration, monitoring, logs
- Tools to back up and sync AI assistant settings, and to set up and back up Postgres
- Written guides on agent messaging, orchestration and coordinating across projects
- The Flywheel scripts and its v2 design with three run modes
What isn't there yet
- No installer or setup guide for someone else's machine
- No automated tests
- Scripts assume our folder layout and project names
- Flywheel would need edits before it could run anywhere else
Under the hood
- Plain Bash scripts around git
- NATS, a lightweight message broker, carries agent announcements and keeps history
- Flywheel ran headless AI sessions on a cron schedule, keeping all its state in plain Markdown files
- Patterns were pulled from two larger multi-agent projects where they were first used