Discipline Over Swarms: Making Agents Prove Their Work

Several prominent builders converged this week on one theme: agents do better work when they are made to prove it.

Addy Osmani wrote that giving an agent a way to verify its work “can dramatically improve the quality of output.” With that process in place, he said, Claude Code was able to write more than 90% of its own code. When software is generated faster than people can review it, “the bottleneck shifts to proving” that it works.

Robert C. Martin (Uncle Bob) announced Agentic Discipline 7, writing that “the models have changed.” A single disciplined agent, he says, can now take on big tasks that used to need a swarm-forge harness. The installment covers three big tasks, one deep case study, “a new twist on Crap/Mutate/Dry,” and “a look just above the code.”

@poteto, who works on Grok Bot, said she is working on a new verification skill for pstack.

Matt Pocock offered a concrete guardrail as his prompt of the day: have an agent create a pre-commit hook that blocks files over 5,000 tokens, estimating tokens as bytes divided by four. The hook runs only on staged files, ignores files not authored by agents (such as lockfiles or generated code), and includes explicit instructions in its error message on how to split an oversized file.