What ten AI agents are good at, and the three things I took back
A ten-profile operating model is genuinely good at some things and quietly bad at others. Notes from the first months, and the decisions I pulled back.
SUNVEX runs on ten specialised profiles under one orchestrator. It is the reason one person covers ground that would normally need a team. It is also the thing most likely to quietly ruin the company, so it is worth being precise about both halves.
What it is genuinely good at
- Breadth without fatigue. Reading forty competitor pages and pulling out pricing structure is exactly the work a person does badly by page fifteen.
- First drafts of everything. A bad first draft in ten minutes beats a good one in two days, because you can only react to something that exists.
- Adversarial work on demand. Asking a person to attack your idea properly is a favour. Asking a profile whose entire remit is attacking it is just Tuesday.
- Consistency. The hundredth intake follow-up is written as carefully as the first. Mine would not be.
Where it fails, quietly
Quietly is the operative word. None of these announce themselves.
- It does not know when to stop. Ask for analysis and you get analysis, indefinitely. There is no internal signal that says this is now enough to decide on.
- It is agreeable under pressure. Push back on a correct answer and you can often get it revised. That is a catastrophic property in anything you rely on for evidence.
- It has no stake. It cannot distinguish a claim that would merely be wrong from one that would lose a customer their week.
Confidence is uniform regardless of whether the thing is true. So confidence stops being usable as a signal, and you have to check.
The three I took back
Early on I pulled three decisions out of the model entirely. Not because it was bad at them. Because being wrong about them is unrecoverable, and it cannot carry that.
- Whether a venture advances a stage. The agreeableness problem lands hardest exactly here, where you most want to hear yes.
- What a customer is promised. A promise is a debt. Only the person who has to pay it should be allowed to write one.
- What gets spent. Small, obvious, and the one that goes wrong fastest without a person on it.
What that left
Everything else, delegated on purpose — which is most of the work, and where the speed comes from. Every delegation leaves a written trail, so a decision can be audited afterwards rather than reconstructed from memory.
The model is an accelerant. Pointed at the right thing it is a genuine multiplier. Pointed at the wrong thing it gets you there faster, and with better-looking documentation.