Agentic Swarm Simulator

Inspiration

I served my app Overturn from the same commit in two places. On Vercel its menu worked; as plain files, About returned 404. I knew the intended route, so this was easy for me to miss. I wanted a first visitor that could show the exact point where the app stopped working and hand back a check I could keep.

What it does

Give Agentic Swarm Simulator a public repository at an exact commit and one plain-language goal. Nemotron visitors take read-only browser steps while the studio saves their screens, actions and claims. When a visit stops on a confirmed error page, the repair loop returns a Playwright check. After a patch, it runs the same check and sends a fresh visitor at the same goal. It credits a repair only when the check was red for that error, goes green after the patch, and the second visit reaches the goal. The hosted app lets judges replay recorded visits and inspect the test, patch and receipt; new repair runs start from the source with the builder's own Token Factory key.

How we built it

Python and Playwright run isolated browser visits and turn a stopped route into an executable check. NVIDIA Nemotron 3 Nano chooses each short browser action through Nebius Token Factory. The broader studio also uses Nemotron 3.5 Super VL to inspect claimed completion against a screenshot; deterministic task rules and retained screens keep a model claim from becoming a result by itself. The public service runs on Google Cloud Run. Recorded playback makes no model call; the hosted Ask Nemotron action is a separate guarded live route.

Challenges we ran into

Our first Overturn run said the check was red before and after a fix. It was red for the wrong reason: CSS rendered the menu as ABOUT, while the link's accessible name was About, and the generated selector missed it on both versions. We repaired the generator and required the original red check to be caused by an error page.

Accomplishments

The public browser bundle retains 12 saved real-app runs across 8 distinct apps: 7 external projects and our own Overturn, with 32 visitor trajectories. Three controlled demo/sample runs are separate. In the complete owned Overturn repair case, a Nemotron Nano visitor hit an About 404 on a plain static host. A five-file link patch made the returned navigation check pass; a fresh recorded visitor reached the same goal. The original Vercel deployment also passes, so this is a host-specific reproduction. The 13 Token Factory replies report the requested model ID and usage: 16,317 prompt and 617 completion tokens, about $0.00113 at listed rates, not an invoice. This is one repair case, not a success rate.

What we learned

A model visitor can find a broken path, but it can also misunderstand a working page. The useful result is the saved screen plus a check tied to the actual failure, followed by a fresh visit. The checker itself needs the same scrutiny as the app it checks.

What is next

Run more apps using their own build and server instructions, not only static files. Let a builder start a new bounded run from the hosted UI once a durable provider-level spend stop exists. Compare model visits with consenting human first visits before making claims about usability.

Live demo: https://agentic-swarm-studio-568004190078.us-central1.run.app/ Public source: https://github.com/Morkeeth/agentic-swarm-studio

Built With

• css
• html
• javascript
• nebius-token-factory
• nvidia-nemotron-nano
• playwright
• python