What it is
PlatAtlas is the platform I spend most of my time building. It takes the work that agents and machines do, the kind of work that usually vanishes into a chat log or a robot’s memory, and turns it into a record a third party can actually check.
Each organization gets its own console. The console maps the org’s work as an atlas: the actors, the actions, the traces. Every action is signed, so the record is attributable and tamper-evident, and it is replayable, so verifying what happened never requires taking the operator’s word for it. That last property is the whole point, and it is the same thread that runs through my essay, Facts a third party can check.
How it fits together
PlatAtlas is a platform, not a single app, and it has an open-source edition underneath it:
- The platform is the hosted product: per-org consoles, signed action traces, a fleet view for machines, and cost and capability panels, all running on Cloudflare’s edge (Workers, D1, R2).
- OpenCastor is its open-source robotics edition. That is where the ROBOT.md manifest, the RCAN wire protocol, and the Robot Registry Foundation live, including the first five EU AI Act-aligned compliance endpoints.
So a farm running robots and a team running agents use the same underlying idea: describe the work declaratively, sign what actually happens, and keep a record anyone can replay. Agriculture is the first vertical I have taken end to end; the platform itself is not ag-specific.
Why it matters
Most tools that claim to make AI accountable show you a dashboard. A dashboard tells you what a system says it did. PlatAtlas is built so the record stands on signed facts instead, the kind a regulator, an insurer, or a small offline model beholden to no one could re-check for themselves. It is a prototyping-heavy project, built by pairing with Claude Code, and it is where the accountability ideas I care about stop being an argument and start being infrastructure other people can use.