When an AI agent disappears, and how Stevie Bennett replaced it
Stevie Bennett's chief architect agent vanished overnight when its file corrupted. What went missing, the handover she ran to replace it, and what any team running AI agents should write down before it happens to them.
Stevie Bennett runs Helmworks with a team of AI agents. One of them was the chief architect: the agent that ran the operation alongside her, kept the coding agents focused and knew the project's standards. Then, one week, he was gone.
"Not metaphorically. Not in the 'he got poached by another startup' sense. His file corrupted. Completely unrecoverable."
She wrote it up on HackerNoon as "My AI Employee Died. Replacing Him Was a Nightmare.", and on Helmworks as So... The Chief Architect Vanished. This page is what that week teaches anyone who runs real work on agents, because sooner or later one of yours will go too.
What actually went missing
He had been with her from the start. In her words: "My right-hand man. My consigliere." He knew how to manage the mini-Claudes, the coding workers underneath him, and "The mini-Claudes liked him." He understood the tone, the standards and the rhythms of the place.
"We had developed little shared jokes and working patterns. We had shorthand."
That is the part nobody budgets for. Starting a new agent is easy. Starting one that knows your project, your standards and your shorthand is a hiring problem, and she treated it as one.
The handover she ran
"It was like onboarding a brand new senior architect into a real company from scratch."
She did what a decent company does when someone senior leaves, in three steps.
| Step | What she did | The human equivalent |
|---|---|---|
| The written standards | Handed over architectural and coding standards, design decisions, operational notes, edge cases, gotchas, governance constraints, patterns and anti-patterns: "Probably forty or fifty separate artefacts at this point." | The wiki, the decision log, the onboarding pack |
| A briefing from a colleague | Asked the synthetic marketing manager for "a briefing pack for the new architect" | A peer who sits the new hire down |
| The history | Gave him the whole conversation trail between the old architect and the coding workers | The leaver's email chains, chat history, commits and meeting notes |
Her instruction with the documents: "Read all of this ... And don't just read it. Absorb it. There will be a test later." And with the history, basically: "Here. This is how the old guy thought."
Then he behaved exactly like a new senior hire
First he tried to impress her. He suggested running the unit tests, verifying the commits and making sure the developers followed process. All of which was already happening. She found herself defending the agent that did the engineering:
"No, no, no ... He already does all of that automatically. He runs every test before every commit. He writes new tests proactively. Don't treat him like an idiot. He understands this project. You're the new person here."
Which is when it landed:
"I realised I wasn't managing tools anymore. I was managing personalities."
He adjusted. Then he did the other thing new senior people do, and started showing off: "Producing impressive-looking but strategically irrelevant ideas because he wanted to establish credibility." Her verdict: "Again, exactly like real life."

What fixed it
Not more documents. Not a sterner prompt. She stopped trying to turn him into a copy of the old architect and gave him something of his own.
"Right. New project. New architecture problem. Go solve this."
"Trying to force him directly into the shape of the old architect failed. Giving him ownership worked."
Anyone who has managed people will recognise that. The surprise is that it holds for agents too.
What to take from it if you run agents
Each of these comes straight from what she did that week.
- Treat an agent role as a role, not a prompt. If the only record of how the job is done lives in one agent's history, you have a single point of failure with a personality.
- Write it down before you need it. The forty or fifty artefacts are what made a replacement possible at all. Without them she would have been rebuilding the job from memory.
- Hand over the history as well as the rules. The standards say what. The trail says how the last one thought.
- Expect the newcomer to over-check and show off. Stop it re-auditing work that is already being done properly. That is your job, not its.
- Give the replacement its own work. Ownership got further than imitation.
The wider operating rules, the written jobs and the independent checks, are on how she runs a company with AI agents.
The bit no vendor guide mentions
The next morning she missed him.
"Not the functionality. Not the output. Not the productivity. The personality."
"Which is, objectively speaking, completely insane."
It is also useful information. If losing an agent feels like losing a colleague, it was doing a colleague's job, and that job deserves the same handover you'd give a person. The full story, told better, is on Fast Track Femme.
If your team runs on agents
Helmworks sets this up for other teams: the written jobs, the standards a replacement reads on day one, and handovers that don't depend on any single agent's memory. Talk to Stevie about yours.
Stevie writes about the synthetic company as it happens, along with everything else, on Fast Track Femme.