Baton

AI agents remember what they said. Baton helps you see what they did.

Every agent user knows the moment. A session gets interrupted, the context is gone, and you are left with two bad options: start the job over, or trust the agent's own retelling. Baton gives you a third option. Give it the record of the run and it checks every claim against the proof, then hands you what completed, what failed, what is waiting on you, and what comes next, each claim linked to its source.

Baton looks at a session after the fact. It does not watch your agent while it works, and there is no background watcher. It reads one finished or interrupted run: what the agent said and did, what it created or changed, proof of what was sent or delivered, and what a human was asked and decided. The demo replays it in three panels: what the agent said, what the proof shows, what changed.

The demo ships four cases. Three are proof cases: a job that never ran at all, work that finished but was never reported, and a message accepted but never confirmed delivered. The fourth is a real run: the session that fixed and shipped this demo itself, recovered from its own history, with that history published alongside it so anyone can repeat the check. A short demo shows it picking up exactly one next step from where the last session stopped, and checking that it got done.

The reading is done by NVIDIA Nemotron-3-Super-120B, running through Nebius Token Factory. If no AI is available, a built-in set of fixed rules does the same job. The AI's private reasoning is never stored, shown, or passed on, and the fixed checks run before any AI is involved, so the AI never grades its own homework. In our own 12-scenario test, Baton got every verdict, next step, and source right, 12 out of 12. Asking an AI to summarize the same sessions got only 4 out of 12 right.

Baton also includes a simple guide that teaches your agent to keep a clean record as it works, so future sessions can be checked the same way. Adapters translate other agents' logs into the same format, each from a single example session. If your agent keeps no record at all, there is nothing to recover afterwards.

Baton is free, open source software under the MIT licence. The repo is public, the demo is live, and the build is the entry for the Best Apps and Agents track.

Demo: https://baton-web-lovat.vercel.app/ Repo: https://github.com/Ized09/baton