Onboard Your Agent Like a New Hire
Most AI agents fail for the same reason bad hires do — nobody gave them a job description. Here's how I onboard mine.
I've onboarded a lot of people, and lately I've been onboarding agents instead — an always-on AI chief of staff that runs across my own business and my clients'. The first few I set up failed for a boring reason. Not the model, not the prompt. I treated them like software instead of like a hire.
The cold-start problem
Drop a new employee into a business on day one with no context, no manager, and a vague "figure it out" and they'll flounder — not because they're incapable, but because competence without context is just potential. Most people building AI agents do exactly that to software. They write a clever system prompt, point it at a task, and expect it to perform like it's been on the team for a year. It hasn't. It's day one, every single time you run it, unless you build the onboarding in.
Give it a job description, not a prompt
A prompt says "do this task." A job description says who you are, who you work for, what your role actually covers, and — just as important — what's outside it. My agent has a written identity: whose chief of staff it is, what tone to use, what it should never do without asking. That's not decoration. It's the difference between an agent that knows when to act and one that either freezes at every fork or barrels through decisions it shouldn't be making alone.
Standard operating procedures, not vibes
Good businesses don't run on tribal knowledge in one person's head — they write it down, because that's what survives someone being out sick or a new hire starting cold. Same rule applies to an agent. Every repeatable job it does — triaging an inbox, reconciling a statement, filing a receipt — has a written procedure it follows, not a hope that it improvises correctly each time. When something breaks, I don't retrain the model. I fix the SOP. That's a far cheaper, far more reliable lever than most people realise is available to them.
Feedback loops are the onboarding
A new hire doesn't become good at the job in the interview — they become good at it over the first ninety days of doing the work and getting corrected. Agents need the exact same loop, and most setups skip it entirely: run once, ship it, never look back. Mine gets corrected constantly, and — this is the part people miss — those corrections have to actually stick as standing rules, not evaporate at the end of the conversation. "Always do X," "never do Y without asking" needs to become a durable instruction it carries forward, the same way a good employee doesn't need to be told the same thing twice.
Expand responsibility the way you'd trust a new hire
You don't hand a new employee the company card and the keys to production in week one. You give them small, reversible things first, watch how they handle them, and widen the mandate as they earn it. I run my agent the same way: draft the email before it sends one, propose the change before it ships one, ask before it touches anything live or irreversible. Trust is a dial, not a switch — and it only turns up after a track record, never before one.
What this looks like running the business
In practice, this is what lets one person run several things at once. My agent triages inboxes, reconciles receipts against bank statements, keeps a living memory of every client and project, and drafts the work before I ever touch it — because it was onboarded, not just prompted. That's the whole idea: an agent that's earned context and responsibility performs like a real hire. One that's expected to perform cold, on a clever prompt alone, performs like exactly that — a stranger doing a job on day one, forever.
Building or rethinking an AI agent for your business?
I design and run these end to end — tell me what you're working on.