Designing agents is mostly job design (and a little prompt writing)

People ask how I “write system prompts.” As if the secret is a paragraph with better adjectives.

I do write prompts. I also design jobs. The prompt is the job description. If the job is mush, the prompt will be mush with nicer formatting.

I already wrote about real security vs security theater on this bench: humans own dangerous verbs, confirm before mutate, read the diff. This note is the upstream half: how the work gets shaped before anyone touches a repo.

Start with the work, not the model

Before I care which model is “smartest” this week, I ask:

• What phase are we in? Discover, build, verify, document, ship?
• What would a wrong answer cost?
• Who is allowed to change the outside world?
• What does “done” look like in one plain sentence?

If I cannot answer those, I am not ready for an agent. I am ready for a sticky note.

Give agents lanes, not vibes

My working pattern is a small roster with jobs: keep phases honest, build, write decisions down, hunt defects, own hard edges like sync and local databases. Names are mnemonics. Routing is the control.

One mega-agent that “does everything” is how you get confident nonsense with a wide blast radius. Narrow jobs make review possible. Review is the security feature that still fits in a human brain.

Small tasks – one agent does just fine. I don’t need three agents to write an email! But when you’re building software, it’s a waste. I build a team because then I have agents focused on specific tasks, some of which can easily run in parallel. They aren’t overlapping – they have clear rails for their roles, and when they need to hand off to a different agent.

Prompts as contracts

A useful agent brief includes:

• Role in one line (“you research and cite; you do not ship”)
• Inputs (paths, constraints, voice)
• Outputs (file locations, format, what “good” means)
• Non-goals (explicit)
• Escalation (when to stop and ask)

That is system-prompt writing, sure. It is also the same craft as writing a crisp ticket. Fancy temperature knobs do not rescue a vague ask.

I prefer models that match the lane (builder vs tester vs researcher) when it matters. I do not treat model shopping as the product. The product is the workflow that still works when the leaderboard rotates.

Phase gates beat pep talks

Discovery before dig. Quality before “ship it.” Dependency hygiene before feature nostalgia. I write gates down because future-me will try to skip them while feeling productive.

Agents are excellent at helping me skip. That is why the gate is part of the design. “Later me” has to argue with “Current me” and my knowledge of where I’m likely to try and fudge a gating task. If I have a document in front of me and an agent telling me I have to sign off before we can close a wave, I’m going to sigh and get it done.

What I am not selling

I’m not selling a universal agent OS. I’m not claiming my roster is the One True Org Chart. I’m saying: if you design agents like you design jobs (with blast radius, handoffs, and adults on publish), you get less theater and more instruments.

If your prompt starts with “you are a genius who can do anything,” rewrite the job.

More soon if I learn something rude from the next wave of tooling. I probably will.

~ Trish