Most organisations that have deployed AI agents did not build an identity system for them. They reused the ones already lying around: a service account with broad rights, an API key in a secrets store, or a delegated token borrowed from the human who triggered the workflow. It works, in the narrow sense that the agent functions. It fails in every sense that matters when something goes wrong or when an auditor asks who did what.
The result is a class of workload that behaves like a user but does not follow any of the rules built for users. It authenticates continuously, acts at machine speed, may chain through other services, and often has no manager to approve its access or review it quarterly. Non-human identity was already the fastest-growing and least-inventoried category in enterprise IAM. Agents compress the problem into a timescale where manual governance cannot keep up.
The audit trail stops being an audit trail
Borrowed credentials break attribution. If an agent acts with a human’s token, the logs record the human. If several agents share a service account, the logs record the account. In both cases the organisation loses the ability to answer the one question that matters after an incident: which process, under whose authority, did this. Fixing that is not a logging problem; it is an identity design problem, and it has to be solved before scale, not after.
Scope, lifetime and delegation
Three properties do most of the work. First, scope: an agent should hold the narrowest permissions needed for its task, not the permissions of the service it lives in. Second, lifetime: credentials should be short-lived and issued per task or per session, so a leaked credential is worthless within minutes. Third, explicit delegation: when an agent acts for a person or another system, that relationship should be recorded as a delegation with limits attached, not implied by a shared secret.
Discovery before policy
You cannot govern what you have not found. Agents built by business units on cloud platforms, low-code tools and third-party automation services rarely pass through central provisioning. Inventory work has to include cloud workload identities, CI/CD service accounts, secrets in repositories, and the OAuth grants that connect agents to SaaS applications. Only when the estate is visible does policy have something to bind to.
Treat agent onboarding as an entitlement decision
The organisations doing this well have folded agent registration into the same workflow as human joiner-mover-leaver. An agent gets an owner, a purpose, a data classification, an expiry date and a review cadence. Unused agents are removed on a schedule. That administrative discipline sounds unglamorous next to agent frameworks and orchestration tooling, but it is what makes autonomous deployment defensible to a risk committee.
Runtime matters as much as provisioning
Provisioning controls what an agent may attempt. Runtime controls what it actually does. Behavioural baselines, anomaly detection on API call patterns, and policy enforcement at the point of access catch the agent that has been manipulated through prompt injection or that has drifted outside its intended task. Static permissions will not contain an agent that has been talked into doing something new.