Thorsten Ball's agent rules: ask for proof, review almost nothing
Thorsten Ball, an engineer at Amp, posted the list of rules that had been going out to friends in DMs, titled "as of October 2026". It took about 1,900 likes on Sunday, and Mario Zechner, who wrote the pi agent, called it "recommended reading. i don't agree with everything, but it's a good data point." It's worth reading because almost every rule is something you can test this week.
The headline rule: "Ask for irrefutable proof. Biggest game changer." Agents are good at producing proof, meaning end-to-end walkthroughs, demo videos, before and after numbers, so ask for it instead of reading the diff. The suggested line to put in AGENTS.md is "this is not evidence." The matching rule is the one that will make reviewers wince: "Review only code with high blast radius. There are <5 central files -- those I review. Database migrations, yes." Closures versus classes, variable names and small duplication are no longer worth human time.
The others push the same way. Use only the best models, because "if a model is wrong 10% of the time, you don't know which 10%." Make the whole development cycle visible to the agent, because "if the agent can't see something, it makes it up." Don't iterate with the agent, let it build tools to iterate: "Give me 4 theories on why this is slow, then check all four." Test your AGENTS.md files and skills: "Spawn five blind agents, control vs. treatment, and find out." And "Don't wait for your agents. Let them wait on you," with scheduled wake-ups and heartbeats as the feature Ball missed most when it was off.
One rule directly contradicts what Drew Breunig said at AGNTCon the same weekend. Ball: "Stop manually managing context... Compaction works incredibly well... Some threads are 46 messages, some are 3,714." Breunig: delete your context around 200,000 tokens. Both can be right. Amp's compaction may be good enough that a 3,714-message thread is fine there and not in a weaker harness. That is the harness-over-model point again.
The bigger shift is where the human sits. Review moves from reading code to judging evidence, and the trust is built by running experiments, not by reading. Keep the bias in mind: Ball works for a coding-agent company, and Amp's CEO replied that "Amp itself is irrefutable proof that this stuff works." Run the blind-agent test on your own repo before adopting the review rule.
Post: https://x.com/thorstenball/status/2114386050638336082
← Back to all articles
The headline rule: "Ask for irrefutable proof. Biggest game changer." Agents are good at producing proof, meaning end-to-end walkthroughs, demo videos, before and after numbers, so ask for it instead of reading the diff. The suggested line to put in AGENTS.md is "this is not evidence." The matching rule is the one that will make reviewers wince: "Review only code with high blast radius. There are <5 central files -- those I review. Database migrations, yes." Closures versus classes, variable names and small duplication are no longer worth human time.
The others push the same way. Use only the best models, because "if a model is wrong 10% of the time, you don't know which 10%." Make the whole development cycle visible to the agent, because "if the agent can't see something, it makes it up." Don't iterate with the agent, let it build tools to iterate: "Give me 4 theories on why this is slow, then check all four." Test your AGENTS.md files and skills: "Spawn five blind agents, control vs. treatment, and find out." And "Don't wait for your agents. Let them wait on you," with scheduled wake-ups and heartbeats as the feature Ball missed most when it was off.
One rule directly contradicts what Drew Breunig said at AGNTCon the same weekend. Ball: "Stop manually managing context... Compaction works incredibly well... Some threads are 46 messages, some are 3,714." Breunig: delete your context around 200,000 tokens. Both can be right. Amp's compaction may be good enough that a 3,714-message thread is fine there and not in a weaker harness. That is the harness-over-model point again.
The bigger shift is where the human sits. Review moves from reading code to judging evidence, and the trust is built by running experiments, not by reading. Keep the bias in mind: Ball works for a coding-agent company, and Amp's CEO replied that "Amp itself is irrefutable proof that this stuff works." Run the blind-agent test on your own repo before adopting the review rule.
Post: https://x.com/thorstenball/status/2114386050638336082
Comments