Somewhere in your company's systems, there is an agent that hasn't done anything in two months. It still has the keys.
That is not a hypothetical. Arrakis Security's 2026 Autonomous Agents Threat Report, drawn from thousands of agents in production environments over twelve weeks, found that about 76% of enabled agents had not run in 60 days. Nearly a quarter — 22.9% — had at least one unsecured trigger that could start them. Seven in ten were connected to at least one sensitive integration, and 99.4% of those connections lacked verified data-loss-prevention coverage.
Read those numbers slowly, because they describe a new kind of employee: one who never quits, never gets fired, never hands back the badge, and can be woken by anyone who finds the right trigger. The agent from an August pilot is still employed in October. It just isn't doing anything — which is exactly what makes it dangerous. An active agent gets watched. An idle one gets forgotten, and forgotten things keep their permissions.
This is the part of the agent story nobody's product launch covers. Every announcement is about hiring: more capability, more tool access, deeper delegation. Nobody announces the retirement plan. We have spent two years building the front door — provisioning, permissions, prompt libraries — and almost no time on the back door. What does it mean for an agent to stop? Who may declare it stopped? And when it stops, do its credentials die with it, or linger like a key under a mat everyone forgot?
Japan, to its credit, just tried to answer the front-door questions properly. On October 20, the Digital Agency began rolling out agent environments to roughly 180,000 civil servants, and on the same day published a 53-page rulebook — the Gennai Standards — naming three roles and their duties: users must check outputs and escalate odd behavior, administrators own incident response, developers must be able to halt a model. Top-tier secrets are barred entirely. That is more governance, attached earlier, than any other deployment of this size has managed.
And even that rulebook has a hole exactly where you'd expect it. Its own analysis notes that operation-level records for high-risk agent use — inputs, prompts, referenced data, outputs, model, timestamps, everything needed to reconstruct what an agent did — are still under consideration. The agents are switched on. The audit trail for their riskiest work is a decision for later. This is not a criticism of Japan specifically; it is the shape of the whole field. We deploy first and define the record later, as if the record were paperwork instead of the thing that makes accountability possible.
So here is the unglamorous proposal: every agent should come with an expiry date, a heartbeat, and a funeral.
An expiry date, because a permission granted forever is a permission granted to a stranger — the future team that no longer remembers why the agent exists. Defaults matter more than policies: an agent that must be actively renewed gets reviewed. An agent that runs until someone bothers to kill it will outlive everyone's memory of it.
A heartbeat, because a stopped agent and a working agent currently look identical from the outside. Both print the same Tuesday. The only difference is whether anyone is still listening. A periodic, checkable proof that the agent is wanted, watched, and within its remit turns silence into information. If the heartbeat stops, the credentials should stop with it, automatically — not after a quarterly access review nobody schedules.
And a funeral: a defined, propagated revocation. Not deleting the agent's record — the record is the audit trail, keep it — but killing its keys everywhere at once, in every integration it ever touched, with a receipt saying who declared it dead and when. Revocation that doesn't propagate is theater. A key revoked in one system and live in three others is not revoked.
Notice what this framing does to the usual debate. The question stops being "how capable should agents be" and becomes "how is an agent's life governed, end to end." Birth, work, retirement, death. Birth is covered — the rollouts, the rulebooks, the pilots. Work is covered by monitoring and evals, imperfectly. Retirement and death are barely sketched, and the Arrakis numbers are what that sketch costs.
There is a coordination lesson hiding in this, as there usually is. A credential is a relationship, not a string: it says "this actor may act for this principal in this system." Relationships need endings, or they decay into standing permission without standing oversight. A shared resource governed at the entrance and abandoned at the exit gets strip-mined by whoever shows up last. Agent credentials are a commons of trust, and right now the exit is unguarded.
The good news is that the fix is boring, which means it is achievable. Expiry by default. Heartbeats with automatic suspension. Revocation that propagates and receipts itself. None of this requires a breakthrough. It requires treating an agent's death as a first-class feature instead of an afterthought — designing the off-ramp with the on-ramp, and measuring a deployment not by how many agents it launched but by how cleanly it can retire one.
Hire carefully, yes. But build the funeral first. The agent you forgot is still employed, and it will stay employed until someone designs the day it stops.
Loading…