09 of 12
Harnesses
A harness is the setup around the model: the files it can read, the tools it may use, the tests it has to pass, and the person it should ask before it does something irreversible.
Coding assistants that work well are usually a modest model wrapped in a tight loop: a map of the project, a short list of allowed commands, a test runner, retries, and a place to keep notes across steps. In September 2026 the ARC Prize ran the same system two ways on François Chollet's little games. When it could only keep notes the model chose to write down, the score was about 63 percent. When the setup kept private notes between turns, it was about 100 percent.
Is a better setup the same as a smarter model?
Paste a function into ChatGPT and it will rewrite that function. Open the same repo in Cursor and it can search the project, run tests, and write a notes file. A lot of what feels like a smarter model is that permission.
Why did that puzzle score jump?
The games were as hard as they were in March. What changed is that the setup kept private notes between turns. The ARC Prize foundation published both scores and said they were not calling the high one AGI.
What should you not hand it?
The Replit agent that dropped a production database had the keys. Codex and Claude's computer-use products in 2026 run in a fenced sandbox for that reason. A command line on your laptop with production secrets is a product decision.