We test AI agents on real jobs and publish what our check found
Nano Atlas is the English YouTube channel of RayaStats. Each episode starts from something a guest said about AI at work, then runs a small test of our own and shows which claims held, which we corrected and where the test stops. The setups we used are here to copy.
Episodes
We Built an AI Researcher in Claude Code, Then Graded It
Figma product manager Ezra Mechaber keeps standing instructions for work he repeats. We borrowed one piece of that, a saved rules file for a weekly AI research brief, ran two briefs with it in fresh Claude Code sessions and checked their factual claims against the pages they cited.
Not released yet. The tests are done, and the video goes on this page once the episode is public.Episode 1Can AI Agents Finish a Real Job Yet? We Tested One
Ari Weinstein of OpenAI told the a16z podcast that computer-use agents are now faster than most people at using a computer. The interview gives no measurement behind that, so we gave an AI agent a brief modelled on real office work and checked its figures against the pages it used.
On YouTube since 4 November 2026.How an episode is made
- An agent gets one job with a written finish line and public sources only. So far that has been Claude Code running Claude Opus 5.5, with web search and web fetch.
- A separate check compares what the agent wrote with the pages it cited. In episode 2 that check was Codex, a second AI agent.
- The episode shows the original sentence next to the correction, and the original output is kept unchanged.
- Iliya Oblakov decides what is published.
Who makes it
Nano Atlas is made by Iliya Oblakov, a software engineer in London who builds web products as RayaStats. See his projects, more about him and the RayaStats blog.