Ponjo: Write together with your AI buddies

Inspiration

Writing with AI today means taking turns: you prompt in a chat, wait, then switch to the doc to see what changed. With people, we stopped working this way long ago: in an online doc, everyone writes at once and talks in comments. I wanted the same with AI agents.

What it does

Ponjo is an open-source editor where you and your AI buddies write the same file at the same time. Comment on a passage and @mention a buddy: It edits only what you asked about and replies in the thread while you keep writing. Every agent's edit can be traced and undone in one click. Every buddy in the web demo runs on NVIDIA Nemotron 3 Super through Nebius Token Factory.

How I built it

Every buddy runs on Nemotron 3 Super through Token Factory, inside an agent harness I built for co-writing. The model gets six tools (read, find, grep, comment, resolve and propose) but never writes the file itself. The harness applies an edit only if the text it quotes appears in the doc exactly once, and otherwise sends it back to fix. Accepted edits then merge into the doc while people keep typing: people's edits by character, a buddy's by line, and where they overlap, the person's text wins.

Challenges I ran into

Nemotron read the wrong file. When a comment meant another doc, it read the one it was in. Listing the project's files in the first message took it from 11 of 27 to 45 of 45 runs reading the right file.
- Nemotron said it did things it never did. It added comments by writing them into the comment file itself, often got the format wrong, and then claimed they were there. Now it calls a comment tool instead: my code saves the comment and tells the model whether it worked, so it can't claim a comment that failed.

Accomplishments that I'm proud of

Real-time co-writing between people and AI: nobody waits for a turn, and a buddy never overwrites what someone typed.
A reliable agent on a mid-size open model: by improving only the harness, Nemotron 3 Super went from 63 to 75 of 81 benchmark runs, and from 11 to 45 of 45 at reading the right file.
A complete, open product: a desktop app plus a web demo anyone can try with no install or account.