AI agents need badges, not API keys
The credential many AI agents run on today is an API key pasted into an environment variable: valid forever, scoped to everything, owned by nobody. That is the same mistake as the SSH key in a drawer, made faster and with more privileges.
Why agents are the sharpest case
An agent is a workload with three habits that make credentials dangerous. It is created and destroyed quickly, often by another agent. It acts on behalf of a person or a service with that principal's authority. And it chains: it calls a tool that calls a service that calls another, so the question "what is this request allowed to do" has to be answered several hops from where the authority started. A static key answers none of that. It says only that whoever holds the key may do whatever the key may do.
What the credential should look like
- Short-lived. Minutes to an hour, so a leaked credential is a small problem and a forgotten agent stops being one on its own.
- Attested. Issued only after the workload proves what and where it is, the same way workload identity does for any service; the standard for the resulting identity is SPIFFE.1
- Delegated, not copied. The agent's authority is a slice of a person's or a service's authority, granted explicitly, and every further hop is a slice of that slice.
- Inventoried and revocable. An agent's credential should appear in the same inventory as every other machine credential, with an owner, an expiry and an audit trail, so that revoking it is an operation rather than an archaeology project.
How trstctl does it
Agent delegation issues a short-lived credential to an AI agent only after the isolated signer verifies a delegation chain: a root principal delegates a narrow slice of authority, every hop must be no broader than its parent, and the signer checks the whole chain before any agent key is generated.2 The signer refuses before key generation if the root anchor is absent, a hop widens authority, attestation fails, or the request falls outside the reachability ceiling, and a lifetime above the sub-hour ceiling is refused before the signer is asked.2 Because the chain check runs in the signer, a compromised control plane cannot slip a widened hop past it. The simpler single-hop case, an agent that just needs an attested workload identity, goes through the agent broker, which uses the same attestation checks as every other workload, and comes back with a full SPIFFE identity you can hand to the receiving service.1
This is new code in an alpha product, and out of the box it issues nothing: every chain-bound request is refused until an operator provisions the signer's trust anchors. The design is deliberately conservative: the ceilings are low, the failures are closed, and the signer keeps its public trust material in its own local store, with no database, message bus or HTTP access.2
The agent delegation page has the chain rules and the signer's refusal conditions; the workload identity page covers the broker.
Sources
- SPIFFE, SPIFFE Overview; trstctl documentation, Workload identity: the AI-agent broker and the signed workload identity disclosure.
- trstctl documentation, Agent delegation: chain-bound identities, the signer's checks, and the sub-hour lifetime ceiling.