Operational infrastructure
Baton
The shared task graph connecting engineering work across an entire repository portfolio — one Portfolio → Project → Task hierarchy that both people and AI agents read and write directly, instead of status living in someone's head or a Slack thread.
Most organizations coordinate cross-repo work by hand: spreadsheets, standups, and whoever remembers to update the ticket. Baton replaces that with a single live graph that every repo, every agent, and every person reads from and writes to — so "what's the state of this work" is always answerable, not reconstructed after the fact.
- A Kanban board tracking work across every connected repository, not just one.
- An MCP server so AI agents can query and update tasks as part of their own tool loop, instead of a human relaying status by hand.
- Bidirectional sync to version-controlled YAML — the task graph is never only a database row.
- Automated review scoring for agent-completed work, so task completion carries an evidence trail, not just a status flip.
Why this exists
Baton wasn't planned as a product. I needed AI agents to be able to coordinate with each other on real work while I could still see exactly what was happening — and I couldn't find a tool that did that. So I built one, and it evolved into the infrastructure it is today: not designed top-down, but grown one real coordination problem at a time.
Real usage, not a demo
- In continuous, daily use since March 2026 — five months and counting.
- 173 pull requests merged and 280 commits on the flagship repo alone.
- 23 actively connected repositories across the portfolio, each reading and writing to the same task graph.
- 4 AI agent identities — Claude (Anthropic), Codex (OpenAI), Gemini (Google), and Grok (xAI) — completing real tasks through the same interface, not a single-vendor integration.
Where this goes
AI agents already open real pull requests and close out real tasks inside Baton today. The industry-wide problem is trust: knowing whether a completed task actually did what it claims, without a human re-checking every line. Evidence-backed task closeout is live now — an agent can't mark work done without recording what changed and how it was verified. A judge-panel model-quality system that scores completed work against a rubric, instead of taking a status flip on faith, shipped its core scoring loop this month, and extending it further is next. Multi-agent software delivery doesn't get safe to delegate until verification is structural, not a courtesy.
Team
Built and operated by Jurie Smit, alongside Mystira and the rest of an actively developed, multi-repo portfolio Baton coordinates work across every day.
Funding
Baton is raising a pre-seed round. If evidence-backed, multi-agent-verifiable work coordination is something you want to back, get in touch.