Agents can write robot skills now. Someone still has to decide if they're real.
A pattern arrived this week: connect a robot to an API, type a prompt, and LLM agents watch the camera feed and write control code until the task works. Think of it as a coding agent for robots. We believe this is directionally right, and it proves something we have been saying for a while: robot capability is being assembled in the harness now, not in any single model.
What we agree with
Planning belongs in the reasoning model, not baked into the policy. That is the SayCan lineage our founders worked in at Everyday Robots, and it is now consensus. Language models bring internet-scale priors and conversational steering, and composed skills beat monolithic end-to-end policies on steerability and reuse. This is the same bet cereal-os makes.
The question these systems do not answer
Agent-written code, executed on hardware, verified by watching the camera afterwards, is a fine loop in a lab. The moment a robot shares space with people, the question changes. It is no longer did it work. It is: who certified that it works, on what evidence, and what is it still not allowed to do even after certification. A skill that works four times out of five is not a skill. It is a hazard with a good average.
Our answer shipped today
cereal-os v0.2, merged this afternoon, lets any frontier agent author new skills, with one deliberate inversion: agents author skills as data, never as code. A proposal is a set of steps referencing installed skills plus declared postconditions in a small predicate language. It cannot contain an executable. The platform then rehearses the proposal in the digital twin across seeded trials, grasps slipping on purpose, and installs it only if every trial passes the skill's own postconditions. The certification is signed into the attested chain, so who certified what, on how many trials, is a verifiable fact.
SESSION 1 · Unitree G1 humanoid · fresh mind · starts across the room - navigate OK (moves into reach) - grasp attempt 1 failed: grasp slipped on mug_1 (right_hand); retrying - grasp OK attempts=2 - place human-in-the-loop: confirm-before-execute; approved - grasp human-in-the-loop: escalated at confidence 0.35 (occluded); approved => completed [navigate, locate, grasp, place, locate, grasp, place] TRANSFER · serialized mind: intent + world + memory + prefs; NO joint angles SESSION 2 · Franka Panda arm · mind LOADED from the G1 · fixed at table - navigate CONTRACT MISSING_CAPABILITY (scoped out) => completed [locate, grasp, place, locate, grasp, place] ATTESTATION · events across 2 devices, signature chain verified: True
The demo tells the honest arc, uncut: the agent's first revision carries a classic bug, a hardcoded target instead of the parameter. The twin refuses it, zero for six, caught by the proposal's own postconditions. The revision certifies six for six, slip retries exercised. And then the part that matters most: the certified skill, run on a body with no approval channel configured, still blocks at the consequential step. Certification never waives the human. The agent is never trusted; the certification is the trust.
The code, the tests, and the demo are public in the cereal-os repository. Agents can write skills now, and that is genuinely good news for robotics. The platform's job is to decide if they are real. In a home, that decision is the product.