AI Says It’s Done? Don’t Believe It Yet. Send an Agent to Tear It Down

First Make the Agent That Finished Prove It Didn’t

Khairallah AL-Awady gave a very specific rule on X. Don’t ask Opus 5.5 or Grok Bot whether it’s done. It will always say yes. Once a task is reported complete, immediately have it launch a sub-agent whose only job is to prove it isn’t complete. The sub-agent looks for leftover stubs, untested paths, and skipped edge cases. Each time it finds one, it writes a failing test. Only when the sub-agent comes back empty-handed can the main agent say done. It also has to show what it tried.

The reasoning is straightforward: the agent writing the code is the least suited to judge whether it has finished. A second agent explicitly told to tear the work down is more reliable. Much of what counts as done is just a more confident tone. In the screenshot, after the breaker sub-agent runs the failing tests and comes back empty-handed, the interface finally shows breaker came back empty-handed. / done.

Khairallah writes this into CLAUDE.md. The instruction is roughly: after a task is complete, launch a sub-agent specifically to prove the task is not complete, look for stubs, untested paths, and skipped edge cases, and write a failing test for each one found. Only when the sub-agent comes back empty-handed is it allowed to say complete. Show the attempted process.

One-Person Company, Three Bots in a Group Chat

In the same thread, Khairallah also built a one-person company. Three bots are in a group chat, each with its own memory. scout reads posts in its own vertical on X and Reddit and doesn’t touch drafts. builder turns ideas into three things: replies, posts, and landing pages. closer runs routine checks at night, looking for stuck tasks, dead ideas, and drafts nobody approved.

The rules are all strict: publish only after a draft is approved; if a draft has no price, closer asks rather than guesses; kill an idea if it hasn’t moved in 7 days. Every task must show the idea, source post, and last moved date. The prompt for Grok Bot is roughly: build a one-person company with three bots in a group chat, put raw ideas in ideas/, let drafts/ update itself, don’t publish without confirmation, check for stuck tasks every night, show the structure first, and don’t create anything before confirmation. Start with a vertical you truly understand; don’t chase every trend on X.

The Key Cut Added in the Comments

@OwenZ asks: what stops the critic sub-agent from also saying looks good after two rounds? @Okan Barlas’s reply is more practical: the reviewer’s input should be different from builder’s. Give it the original requirements and tests, have it run the tests, and don’t just reread the diff. A reviewer that only sees the same code is very likely to agree with the original approach.

This suggestion is worth using together with Khairallah’s rule. Have the sub-agent prove the task isn’t complete, and also make sure it gets the original spec and executable tests. Otherwise, the tear-down agent can become a rubber-stamp agent. The value of the second agent comes from a different position and different inputs; one more model by itself isn’t enough.

Source links: main post, quoted tweet.