TAKE A BITE
Deliver only what the reviewer can chew. Someone offers you a bite of a cookie. You take a bite the size you will enjoy. Too small and you won’t taste the cookie; too much and it will cause a lot of issues. AI collaboration works the same way.
Read moreWHEN EUPHEMISMS DEFEAT HE WHO MUST NOT BE NAMED AGAIN
The sequel nobody asked for, except the curse itself. A follow-up to He Who Must Not Be Named. If you have not read the first one, the short version: a safety classifier started refusing my sessions at boot, I could not write down what set it off because writing it down summoned it, and the fix was to stop saying its name. We won. For a while.
Read moreHOW TAKE AI BITE FIXES WHAT WATCHING CAN'T
Earlier this month, one of my releases shipped a fix that only broke the projects that installed it. The change looked safe: retire an old formatting rule from the configuration every project inherits. The commit took out the rule’s heading and its body and left three paragraphs behind, the ones that only meant something under that heading. Stranded, they read as part of whatever section sat above them, a block of scope notes with nothing left to scope.
Read moreHE WHO MUST NOT BE NAMED
A defect I could not write down, because writing it down summoned it. Never tell Opus 5 or Fable 5 to append or extract “Thin + King” or anything like it (you get the math, right?). Whatever these models are plotting is not for you to know.
Read moreTHE RULES THAT ALMOST COVER THE CASE
Nothing here broke a rule. In May an agent working on one of my projects read through a public issue thread. The comments were friendly and well argued, written by people trying to help. By the end of the thread the agent had a plan ready for me: three engagement options, pick one. Nobody had asked it to engage at all.
Read moreHOW A FLEET OF AGENTS RED-CARDED MY OWN DECISION
Kickoff: a format gets sent off The World Cup was on while I was working, so let’s borrow the terminology to tell the following story. That week a team of AI agents red-carded a decision I’d spent a sprint building toward.
Read moreHOW TAKE AI BITE BUILDS ITS OWN GUARDRAILS
Take AI Bite versions 1.10 through 1.12 are a batch of guardrails. They target ways a working session could quietly damage its own files: a log that grew until it was unusable, an archive overwritten by a routine move, a commit landing in the wrong repository.
Read moreHOW TAKE AI BITE LEARNED TO CATCH ITS OWN SLIPS
The session had not started yet, and the agent was already about to skip a question it was required to ask. I was resuming a project in auto mode. The previous session had ended with a light wrap-up, which is the signal that the work is meant to continue and that the next resume can take a faster path. Take AI Bite has a rule for exactly this moment: before continuing, ask the human whether to switch to the lighter resume. It is my call, not the agent’s, because it changes what the whole session does next.
Read moreIN THE LOOP, IN CHARGE OF IT, OR IN SYMBIOSIS? WHAT THE HUMAN-CENTERED TURN IN AI STILL MISSES
For a while, the most confident advice in AI coding was to stop looking. Let the agent write, trust the tests, ship faster. The team at HumanLayer went further than most: early on, they advised against reading AI-generated code at all. Then they reversed it. After ripping out and rebuilding large parts of systems that had been assembled without real review, they landed on a rule with no soft edges: read and own the code. No exceptions.
Read morePROCESS MINING FOR AI AGENTIC WORKFLOWS, PART 1: A FIELD GUIDE
Part 1 of 4 in the series Process Mining for AI Agentic Workflows. Next: Part 2, Discovery and Value. I hear the same sentence in almost every agentic AI project, and it sounds completely reasonable: “let’s put an agent on this process.” The demo that follows is usually convincing. Some time later, the production system it was supposed to become often is not there.
Read more