What is agent onboarding?

The Two Meanings, Kept apart

It is worth separating the two cleanly, because conflating them is the most common confusion in this area.

Identity provisioning answers: who is this agent, and what is it permitted to touch? It provisions a verifiable identity, attaches permissions and scopes, and creates the audit linkage so the agent's actions can be attributed to it rather than to whoever invoked it. This matured fast in 2026 — multiple major identity vendors reached general availability with agent-identity products, because the security problem is pressing and well understood: an agent with no identity of its own is an attribution and access disaster.

Readiness onboarding answers a question provisioning cannot reach: does this agent actually understand the job it is about to do? Not what it may access — what it knows, what it has been told about how things work here, and whether it is genuinely capable of the specific task. A perfectly provisioned agent, correctly identified and tightly scoped, can still be entirely unprepared. Its credentials say who it is. They say nothing about whether it should be trusted to proceed.

The industry has noticed the edge of this gap. One identity vendor puts it plainly: a verified agent can still operate on uncertified data. That is the seam. Identity stops at the point where readiness has to begin, and mostly nothing picks up there.

Why a System Prompt is not Onboarding

The usual answer to readiness — such as it is — is the system prompt. Write the agent some instructions, paste in a few facts, and consider it briefed. This is to onboarding what handing a new hire a one-page memo on their first morning would be: better than nothing, and nowhere near enough.

A system prompt has three structural problems as an onboarding mechanism. It is a guess: written in advance by someone imagining what the agent will need, rather than discovered from what the agent actually lacks. It is static: it does not get sharper as the agent encounters the limits of what it was told. And it is unverified: nothing in writing a prompt confirms that the agent understood it, or that the agent can in fact do the task the prompt describes. You find out whether the briefing worked by watching what the agent does in production, which is the most expensive possible place to discover it didn't.

Real onboarding, for a person or an agent, is not the act of issuing instructions. It is the process of closing the gap between what the worker knows and what the work requires — and confirming the gap is closed before the work that matters begins.

What Readiness Onboarding Actually Involves

Stripped to its concept, preparing an agent for real work has three movements, none of which a credential or a prompt provides.

The first is finding the gaps. Before an agent acts, the useful question is not "what shall we tell it?" but "what does it not yet know that this task requires?" The gaps are specific to the task and the organisation, and they are easier to surface by examining the job against what the agent has available than by guessing in advance.

The second is filling them deliberately. The missing knowledge — the constraint nobody wrote down, the context that lives in someone's head, the decision that explains why things are done this way — has to be supplied by the people who hold it, and recorded where it can be reused, not re-explained in a disposable session each time.

The third is verifying capability, not just knowledge. Knowing things is necessary but not sufficient. The final movement confirms that the agent can actually perform the task, not merely that it has been told about it — the difference between a briefed worker and a ready one.

The shape is the same one any careful manager uses with a capable new hire on a consequential task: work out what they're missing, supply it, and check they can actually do the thing before the stakes are real. Concretely: an agent asked to handle supplier disputes might draft a fluent, reasonable-looking response and have no idea that this category of supplier is contractually entitled to a longer remediation window than the email it just wrote allows. Finding the gap means noticing that the task depends on knowledge the agent doesn't have; filling it means the person who knows the remediation terms records them where the agent can use them; verifying means running the agent against a few realistic disputes and confirming it now handles them correctly before it touches a live one. Onboarding is that, applied to an agent, and done deliberately rather than left to chance.

Why This is the Harder Half, and the More Valuable One

Identity provisioning is further along for a straightforward reason: the failure it prevents is loud. An unidentified agent that breaches access trips alarms, shows up in audits, and maps onto security problems enterprises already know how to take seriously.

The failure that readiness onboarding prevents is quiet. An agent that is perfectly authorised and badly prepared does not breach anything. It does permitted work from an impoverished understanding, and produces results that are wrong in ways nobody notices until later. There is no alarm for an agent that stayed inside its permissions and got the substance wrong. That is precisely why the readiness half has been under-built: the cost of skipping it is real but deferred, while the cost of skipping identity is immediate and visible.

Both belong under onboarding, even though the mature half is more precisely called provisioning. An organisation serious about deploying agents on real work needs both — the credential that says who the agent is, and the preparation that determines whether it should be trusted to act.