Roughly, the more tokens you throw at a coding problem, the better the result. Boris calls this test time compute.
- Using separate context windows makes the result even better — this is what makes subagents work, and why one agent can cause bugs and another (using the same exact model) can find them
- Similar to engineering teams: if Boris causes a bug, his coworker reviewing the code might find it more reliably than he can
- In the limit, agents will probably write perfect bug-free code — until then, multiple uncorrelated context windows tends to be a good approach