Jared Zoneraich, Builder in Residence at Cognition (the team that built Devin), argues that most teams fail at their first agent because they overbuild it. The fix is counterintuitive: let the model run before adding rules, prioritize better tools over better prompts, and only write evals after you have shipped something real.
The demo at 25:39 is the reason to watch this in full. One manager agent breaks a task into parts, spins up 10 cloud agents in parallel, and decides autonomously what requires human review before returning results. The discussion at 16:06 on what cloud agents can do that local agents cannot is a prerequisite for understanding why that architecture works.
The throughline here is discipline against complexity. Zoneraich is not selling a framework. He is describing a sequence: simple prompt, correct tools, self-testing, then orchestration. Teams that skip to orchestration first are building on sand. The conversation at 28:03, covering how agents prove their own work, is where the practical methodology gets specific.
[WATCH ON YOUTUBE →]