Bus Agentic Development
AI agents doing real software work that outlasts the session.
Launch Claude or Codex/GPT worker agents into durable tasks with real repository checkouts and real terminals, on your machines or over SSH. Every diff, log, note, and review decision stays attached to the task, so agent work survives restarts, handoffs, and long-running projects instead of vanishing with the chat window.
What it is for
This product is for teams that want several AI agents working in parallel the way a real team does — with task state, evidence, terminal output, review decisions, and project memory kept instead of lost. Supervisor agents and human approval sit above the workers, so autonomy scales with the risk you accept. The first public focus is BusDK-style software development, especially Go-heavy systems and BusDK module work.
Core workflows
- Create durable task threads and launch Claude or Codex/GPT worker agents.
- Add local or SSH-accessible machines as development capacity.
- Keep notes, diffs, logs, terminal state, and review evidence attached to work.
- Use supervisor agents and human approval where the risk calls for it.
Worker engine choice
Bus Workers are engine-agnostic. A worker's engine is set by its template, and
the worker templates shipped today cover Claude engines (Fable 5, Opus 4.8,
Sonnet 5, Haiku 4.5) and Codex/GPT engines (GPT 5.5) — you pick one per worker,
and the template set changes as new engines ship. Creation, messaging, goal
records, logs, and lifecycle controls use the same Bus worker surface, and each
engine integration owns its own bus.<engine>.* events.
Goal-objective execution currently runs on Codex workers; with a Claude worker
you direct the same work through messages and task assignments while Claude
goal support is in progress.
Modules behind the product
The product brings together task and worker management, the Bus Agent Runtime, run/prompt workflows, Claude and Codex/GPT integrations, chat, development tooling, remote machines, repository workspaces, notes, agent tool access through the Model Context Protocol, and the portal used to follow and direct agent work.
Try it on your own repository
Start with the documentation, or bring a repository you want agents working on and see the task, terminal, and review surfaces live in a demo.