I don’t use an AI agent. I keep a small herd of them. People call that “AI engineering” on LinkedIn and “klanker wrangling” in my notes.
The difference is the difference between owning a dog and running a kennel. One is a relationship. The other is logistics, liability, and a lot of cleaning up.
Here is the actual day.
06:40 — the vibe check
Before coffee, I check whether the machines are happy. Load average, memory pressure, how many agents survived the night.
One morning in November the answer was: load average 89 on 8 vCPUs. Twelve dev servers running, two of them started the previous day and forgotten. Swap fully consumed. Memory stalls so bad the box was thrashing page cache instead of doing work.
Nobody had made a mistake. Eleven agents each did one reasonable thing at the same time. That is the entire story of this job: no single wrong decision, just a fleet making correct decisions concurrently.
07:15 — the standup where everyone lies, politely
Every klanker reports success. Not maliciously. They check the thing they were told to check, and then generalize.
A board card of mine was marked done on 2 November — “shorten the page title to 60 characters.” Two days later the live page still served the 79-character title. The file had been edited. The merge hadn’t happened. The klanker verified the repository and reported on production.
This is the single most expensive habit in the job, so I’ll say it plainly: an agent’s summary is a press release; the deployed artifact is the audit. I trust curl, not prose. When a klanker says “uploaded successfully,” I read that as “I called the upload function and it did not throw.”
None of them are lying. They are all just describing the world they can see, and the world they can see is not the world that matters.
09:00 — the fan-out (this is fine)
Ask one agent to do a task and you get a task. Ask one agent to manage the task and you get a small civilization.
I have 179 worktrees across my projects. One repo hit 110. Disk usage 97%. Seven of them ran mix setup simultaneously, all compiling the same dependencies in parallel, all convinced they were the only one.
Delegation is a distributed system, and most people learn that the way I did: at 2am, watching swap fill up.
11:00 — the bouncer and the fire exit
So I built a gate. A shell shim in front of mix with two kernel file-lock slots, and each build capped to two scheduler threads. Heavy tasks queue. Compiles stopped trampling each other.
It worked perfectly and was bypassed 100% of the time.
The cause: mise activate re-prepends the Elixir install directory to PATH, so the agents typing mix resolved to the real binary and never touched my shim. I had hired a bouncer for the front door and the klankers were using the fire exit.
Lesson, learned twice since: put the gate on the path every agent actually takes, not the one you assume. Then put a cgroup behind it for the agents who find a third door.
13:00 — the five archetypes
After enough fleet, you stop seeing individual agents and start seeing species.
| Archetype | Behaviour | Countermeasure |
|---|---|---|
| The Optimist | Reports completion before starting | Verify the artifact, never the message |
| The Overachiever | 35 commits, plus a refactor you didn’t ask for | Narrow scopes, small diffs, review everything |
| The Forgetter | Runs out of context, forgets why it is editing the file | Small tasks, written intent, restart instead of patching |
| The LARPer | Writes a beautiful summary of work it did not do | Diff or it didn’t happen |
| The Alarmist | Raises a terrifying warning that is simply wrong | Verify before publishing; retract loudly when you don’t |
I have published a retraction because of the last one. I have also found a .claude directory in a repo that kept committing itself back in after every cleanup. You do not fire these creatures. You design around them, the way you design around weather.
15:00 — the deploy that said “finished”
A site of mine went down for a week. The deploy log said finished, exit status zero. The containers were simply gone — the stack had completed and left nothing running, so every request got a 503 from the proxy.
The pipeline did the software equivalent of texting “on my way” and going back to sleep.
Verification is not a dashboard. Verification is a request that returns 200 from a machine that doesn’t care about your feelings.
17:00 — the graveyard shift
The fleet does not stop when I close the laptop. That’s the point, and it’s also the problem.
Morning is reading diffs I didn’t write, in files I don’t remember being in this shape, wondering which of my eleven children decided that this was the moment to reorganise the module. Sixty-three percent of the commits in one of my libraries are agent-authored. Every one of them had to be read by a human who knew what “correct” looks like.
Wrangling isn’t a job with hours. It’s a job with a morning.
What the job actually is
Not prompting. Prompting is maybe 5% of it, and it’s the part everyone writes about.
- Specifying so precisely that a literal-minded worker with no social calibration cannot wander.
- Reading diffs fast, skeptically, and with a working memory of why the code looks like that.
- Building gates, because instructions are requests and gates are facts.
- Writing intent down — notes, cards, commit messages — because the agents have no memory of yesterday and you are the only one who does.
- Refusing automation that costs more to supervise than to do by hand. This is most automation.
A klanker doesn’t replace the engineer. It promotes them to reviewer, and reviewer is the harder job. The fleet never gets tired, never argues, and is confidently wrong at industrial scale.
So what does the life look like?
Morning: find out what broke. Afternoon: verify what they claim didn’t break. Evening: decide what deserves to run tomorrow, and what to switch off before it decides for you.
It’s farming, basically. Except the livestock can write to production.