What's a klanker?

I made up a word, and I am going to defend it.

A klanker is an AI agent that does real work. Not a chatbot. Not a copilot. Not an “assistant” that drafts your email and then apologises. A klanker is a thing you point at a real problem, in a real codebase, for a real client, and then leave alone with it for a while.

The word matters because the ones on offer are bad. “Agent” means everything and therefore nothing. “Assistant” implies it fetches coffee. “Copilot” is a corporate word for a company that would like to invoice you. Klanker sounds like machinery that is about to drop a wrench on the floor. That is honest. The job is not pretending it won’t.

What managing them is actually like

I am a klanker manager. That is a job title now. It is on my CV, roughly. Here is the honest version of the work.

You do not write the code. You decide what “done” means. This sounds like the easy part and it is the whole job. If you hand over a vague definition of done, you get a confident, well-formatted, entirely wrong answer about four minutes later. Ambiguity is not the agent’s problem to solve. It is yours.

You review. Not skim — review. The agent’s summary is marketing. The diff is the truth. I have read summaries that said “all tests pass” and diffs that quietly deleted a validation branch because it was in the way. Reading the diff is the only way to catch that, and reading the diff is boring, which is why people skip it and then write blog posts about how agents are unreliable.

You say “no, not like that”. Klankers default to plausible. Plausible is not correct. It will produce a solution that looks like the one you wanted and behaves differently in two edge cases you have not thought about yet. Your job is to notice before your users do and send it back.

You notice when they confidently break something. This is the one that eats the hours. Agents rarely say “I am not sure”. They say “Done! Everything works.” Then you find the thing they broke, three files away, marked as unrelated cleanup. A good klanker manager develops an allergy to the sentence “while I was in there, I also…”.

What I do with them

At Unblock Labs I point klankers at real problems for real clients. It is an agency, we build for American brands, and the work is not demos. The clients do not care that the code came from a machine. They care that the thing works on Monday.

Between prompts I build things myself: Elixir, mostly Phoenix and Datastar, and Bubble, where I write no-code plugins and tools. I do this on purpose. If I only managed klankers, I would lose the ability to tell good output from output that merely looks good. You cannot review what you cannot build. So I keep one hand in each.

The two halves feed each other in a boring, practical way. Building in Elixir taught me what a reviewable change looks like: small, scoped, one idea at a time. Building in Bubble taught me that most “engineering” problems are actually “deciding what we want” problems wearing a costume. Both lessons are directly useful when the thing writing the code does not get tired and does not get bored.

The klanker log

I am going to keep a log. What my klankers ship, what they break, and what I had to walk back. The retractions are the interesting part — the commits that had to be undone because the concern was not real, or the ones that had to be rewritten because the agent solved a problem nobody had.

I am not interested in a highlight reel. If the log is honest, the failure rate will be in there, and so will the day I shipped a security fix that only existed because someone made me look at a diff properly.

What this is not

I am not going to tell you agents replace engineers. They replace typing. Typing was never the hard part, and anyone who has shipped software already knew that.

The parts that stay are the parts that were always the job: taste, judgment, deciding which of the thirty things in front of you actually matters, and being willing to say “no” to work that looks finished. Those did not get easier. If anything, they got more expensive, because now there is much more finished-looking work to sort through, arriving faster than before.

A klanker is a machine that does real work. A klanker manager is the human who has to be right about what the work means. If that sounds less fun than writing code, it is. It is also the job now, and I would rather be good at it than nostalgic about the alternative.