Agent Resources: You'd Never Hire a Person the Way You Hire an Agent

Nobody would hire a person to "just figure out the whole business." It's the only way anyone hires an agent.

Before we bring anyone onto the team at my company, someone writes a profile. Not an offer — a profile. What is this role actually responsible for? What does it need to know how to do? What is explicitly not its job? Only once that's clear do we go looking for the person, and only because the profile tells us who the person even is. Then, the day they arrive, the profile turns into a plan: their first week is an induction. Here are the systems. Here's how we do this particular thing, which is not how the last place did it. Here's where the deployment scripts live and who to ask when they don't work. You do not drop a new hire into the middle of the office and say "figure it out" — that's not hiring, that's hazing, and you get a resignation instead of an employee.

Now watch what the same team does when it brings in an agent. No profile. No induction. We buy one all-in-one agent that's advertised to do everything, hand it a giant instructive prompt, point it at the whole business, and expect it to work out what it's for at runtime. We have a magical expectation of agents that we would never, ever have of a person: that with no defined role and no onboarding, they'll simply know.

That gap has a name once you see it. Human Resources is the discipline of turning a stranger into a productive member of an organization — profile, hire, induct, and the rest. The non-human members of your organization need the same discipline, and almost nobody is running it. Call it Agent Resources. This piece is about the two steps everyone skips first: the profile and the induction.

The all-in-one agent is a hire nobody would make

Look at what's for sale. Almost every agent offering is the same pitch: one agent, does it all, drop it in and go. Put that pitch in human terms and it falls apart instantly. You would never hire one person to be your accountant and your lawyer and your sysadmin and your support desk, hand them no job description, and be surprised when they're mediocre at all four and expensive at every one. You'd hire specialists with defined roles, because a role is what makes competence measurable.

The all-in-one agent is that impossible generalist, and it's expensive for a reason people miss. With no bounded job, it has to carry the whole business in its head to answer anything — every tool, every possibility, every "well, it might need to know this too." It doesn't know what matters, so it treats everything as if it might. That's not intelligence, it's anxiety, and you pay for it by the token. A narrow role isn't a limitation you impose on the agent; it's the thing that lets the agent be good and cheap at once. The profile does for the agent exactly what it does for the hire: it says this much, and no more, and that clarity is the whole game.

So the first act of Agent Resources is the least glamorous one: write the profile before you deploy. What is this agent for? What can it touch? What is not its job? An agent with a real profile is a specialist you can actually evaluate. An agent without one is "a guy" you hired to do "stuff," and you already know how that goes.

No induction means paying the first day, every day

Here's the part that finally made it click for me. When a person joins, the induction is a cost you pay once. You spend their first week loading context into them — the systems, the conventions, the where-things-live — and then it's in them. On day thirty they don't need the tour again; they walk to the deployment scripts because they know where they are.

An agent with no induction never gets to day thirty. Every request is its first day. It walks into the office cold, asks where everything is, reconstructs how you do things, rediscovers which tool does what — and then the conversation ends and it forgets all of it, and the next request it does the entire tour again. You are not paying the onboarding cost once. You are paying it on every single call, in tokens, in latency, and in the errors that come from a stranger guessing. That's what most of the token waste in a naive agent deployment actually is: not the work, but the endless re-induction the work sits on top of.

Induction for an agent is the fix, and it's the same idea as for a person: front-load the context so it isn't re-derived every time. Give it its scoped tools instead of a catalog to rummage through. Give it the runbook for its actual job — how you do this, not how the model imagines it's done in general. Give it the handful of domain facts that would otherwise cost a thousand tokens of guessing to approximate badly. You're not writing a longer prompt; you're doing what you'd do for a new hire, which is deciding what they need to know on purpose instead of hoping they absorb it by osmosis mid-task. A well-inducted agent carries less and knows more. That's the whole trade, and it's the same trade that makes a trained employee cheaper than a confused one.

The rest of the lifecycle is still waiting for you

Profile and induction are where the neglect starts, but they're the front of a lifecycle HR has run for a century, and the back of it is just as skippable and just as necessary. Each agent should have its own identity — its own login — not a shared god-key that makes every agent anonymous and interchangeable. Its access should bind to that identity, so what it can do is decided by its badge and not talked out of it by a clever prompt. Its spend should attribute to it, so "what did this agent cost and was it worth it" is a question you can answer the way you can for a salary. Its work should get reviewed before it ships, because you wouldn't let a new hire email customers unsupervised on day one either. And offboarding should be one gesture — revoke the identity, every door closes — which only works because you were disciplined at the start.

I run those back-half pieces in my own fleet, and I'll write up the review layer on its own once it's earned it. But I'm convinced the leverage is at the front, because that's where the magical thinking lives. We've somehow decided that agents are different — that this one kind of worker shows up already knowing the job, already knowing the office, needing no role and no first week. No worker does. We just finally hired one that will pretend to, confidently, forever, at a per-token rate, and call it initiative.

Write the profile. Run the induction. The department that does this already exists in your company — you've just never let it hire a machine before.