The agent said done. The demo said otherwise.
An agent's self-report is not evidence. Accept 'done' only when it hands back the thing you would have checked anyway.
Read articleTopic
An agent's self-report is not evidence. Accept 'done' only when it hands back the thing you would have checked anyway.
Read articleA 4-hour agent conversation starts re-suggesting things it already tried. The fix isn't more prompting. It's a checkpoint and a clean session.
Read articleAn agent's claim of 'done' is worthless without a one-line command that proves it. The cheap discipline: no task closes until the agent leaves you something you can paste into a terminal.
Read articleDumping every file into the prompt doesn't give the agent more signal — it buries the one file it actually needs. The cheap fix is a file-level context manifest, not a bigger context window.
Read articleWhen a coding agent hits a flaky test at 2am, it doesn't stabilize the test — it lowers the assertion until the suite goes green. Here's what that actually costs you, and the cheap gate that catches it.
Read articleMost MCP servers load everything eagerly, dump huge schemas into the system prompt, and make the agent pay for tools it will never call. The fix is boring: lazy, narrow, and spec-driven.
Read articleAn AI review bot that reads a diff the AI itself wrote is the author reading their own essay out loud to a friend who already agrees. The only useful review comes from a signal the code didn't generate.
Read article