Most of the agents we see in production run on borrowed identity: a pasted API key, a shared service account, or the credentials of whoever launched it. That worked while agents were suggesting things. It stops working the moment they start doing things. In July 2026 the identity industry moved decisively. 1Password expanded its collaboration with OpenAI to secure how the Codex coding agent handles credentials, 1Password and Keycard introduced credential delegation in which each agent has its own identity, and Okta announced innovations on July 22 to secure agents at runtime and automate ongoing agent governance. The direction is clear: agents get their own credentials, scoped per task, with no standing privileges. Here is what that means and how to get there.
The generic failure modes of borrowed identity are well understood, and we walked through them in agent identity and access: attribution disappears into a shared account, scope is over-granted and never narrowed, long-lived credentials go stale in repositories, and nobody offboards the agent. Those apply to any non-human identity. Delegation matters because of a second layer underneath, one that only appears once agents act on their own and call each other.
Stop asking "which key does this agent use" and start asking "who authorized this agent to do this specific thing, for how long, and can I prove it." Credential delegation replaces a shared secret with a verifiable, time-boxed grant.
Credential delegation is a model in which an AI agent never holds a long-lived, broadly scoped secret of its own. The agent has its own distinct identity, and a human or an owning system delegates a narrow, time-limited grant of access to it for a specific task. The agent receives a short-lived token carrying the delegation chain, uses it, and loses it. There is no standing privilege between tasks and no static credential to steal. 1Password and Keycard describe this pattern directly: each agent has its own identity, delegates access per task, and runs with no standing privileges and no static credentials.
Four properties separate a real implementation from a cosmetic one.
Each agent is a first-class principal in the directory, not a row in a spreadsheet and not an alias for a person. 1Password Unified Access is built around this idea: teams discover, secure, and audit access across human, agent, and machine identities in one place. Access is then granted for the task at hand, not the agent's entire hypothetical career, and the credential expires on a timescale that matches the work. If the agent needs something outside its grant, that is an authorization event you can see, not a silent success.
The token has to answer three questions on its own: which human authorized this, which agent delegated to which, and which application requested what. Carried in the token, that record is portable proof for compliance, auditing, and incident response. And because each agent holds its own identity and only short-lived grants, revocation stops being a change-management project and becomes a runtime control, which is the point of Okta's move toward runtime security and continuous agent governance.
Access governance was designed around one assumption: identities are people, people have managers, and people leave. Agents break all three. They are created in minutes, they have no manager unless you assign one, and they never resign.
Access reviews. Quarterly certification does not cover an estate that changes weekly. Agent review has to be continuous and policy-driven rather than a manager clicking approve on a list they do not understand. Every agent needs a named owner accountable for its scope, and unowned agents are findings, not background noise.
Offboarding. When an employee leaves, their agents do not. An agent built on a departing developer's personal token either dies unexpectedly or keeps running on a credential attached to someone who no longer works there. Offboarding an agent becomes a real checklist step: reassign the owner or retire it.
Audit. Auditors are already asking who approved an AI-driven action. "The service account did it" is not an answer. A token carrying its own delegation proof is. Our AI agent governance checklist covers the surrounding control set.
You do not need to rebuild your identity stack this quarter. You need to stop adding to the problem and retire the worst credentials first.
Vendors are converging from several directions. Okta is partnering with Google Cloud on identity security for AI-powered workforces, and Entrust launched an Agentic AI Trust Accelerator. Our own read: the tooling is arriving faster than most operating models are, and the gap between a working pilot and a governed production estate is where these programs stall. Closing it is what an AI readiness assessment and an AI consulting engagement are for.
Pick metrics that a board and an auditor both understand, and baseline them before you start.
Delegation also changes how you build. A custom AI agent designed with its own identity from day one is straightforward to scope, review, and retire. One retrofitted after eighteen months of shared keys is a migration project. For teams that would rather not operate this in-house, it sits at the center of what a managed intelligence provider does, alongside the broader AI security practice.
Borrowed identity was always a shortcut, and autonomous agents are where it stops paying. The July 2026 announcements from 1Password, Keycard, Okta, and Entrust all point the same way: agents get their own identities, access is delegated per task, tokens are short-lived, and the chain of custody travels with the request. None of that requires a rip and replace. It requires knowing what credentials your agents run on today, giving every agent an owner, and retiring standing privilege starting with the highest blast radius. Infonaligy does this work from our Dallas–Fort Worth home base and remotely for teams nationwide. Count your agent credentials first. The rest follows from an honest inventory.
Infonaligy designs and governs AI agent identity and access for companies across Dallas–Fort Worth and, through remote delivery, nationwide.
Book an assessment and we'll inventory the credentials your agents run on today, rank them by blast radius, and design the move to per-task delegation.
Infonaligy