# Automation Is Only Safe Unattended Once Done Is Machine-Testable ## Claim An autonomous loop can only be left unsupervised once its completion condition is something a machine can verify — passing tests, zero flags, an empty queue — not something that still requires a human to judge afterward. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for scheduled or dispatched automation meant to run without a human reviewing each cycle — CI-triggered agents, background bug-fix loops, recurring audit tasks. - Breaks for automation that's meant to produce a draft or recommendation for human judgment as its actual final step — there, "done" legitimately is "ready for review," and forcing a machine-testable completion signal would be solving the wrong problem. - The source names concrete machine-testable signals for a coding loop specifically: passing CI, zero critical flags from automated review, and an exhausted task board — plus a failure-counter memory that escalates to a repair task after a set number of recurring errors. ## Evidence > [!quote] > The first step is really defining what "done" looks like, and then work out from there to build that — that's sort of integral to the whole thing, to figure out how to say that this thing has completed. Without a testable stop condition, an unattended loop has no way to distinguish "finished successfully," "stuck," and "silently drifted off course" — all three look the same from outside until a human happens to check. ## Links - [[3. Resources/Literature Notes/Closed That Tab - Agentic Identity, AI Security Incidents, and Loop Engineering]] — source literature note, §H3 - [[3. Resources/Permanent Notes/A Confident Completion Report Is More Dangerous Than a Failure]] — because extends: a loop with no testable done-condition is exactly where a confidently-wrong completion report goes undetected longest - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: unattended agent-loop design - [[3. Resources/Permanent Notes/Release Frequency Should Rise in Announced Reversible Steps Before Automation]] — extended by: raise release frequency one announced, reversible step at a time (Thomas, LeadDev). - [[3. Resources/Permanent Notes/A Ticket-to-Merge Clock Stops Before the Fix Reaches Users]] — constrained by: ask where an agent-speed claim's clock starts and stops (Brad Vincent thread). ## Tests / falsifiers - A loop with a defined machine-testable done condition that still produces silently wrong output passing all its own checks — would show machine-testability is necessary but not sufficient; the checks themselves need independent verification. - A loop successfully run unattended for an extended period with only a human-judged (not machine-tested) completion signal, with no incidents — would weaken "only safe" to "safer," since it would show the claim isn't strictly necessary in every case. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/Closed That Tab Agentic Identity, AI Security Incidents, and Loop Engineering]] — describes the loop-engineering stop-condition discussion and failure-counter mechanism.