A durable-execution platform’s own outages counted against its tenants’ retry budgets and left their runs marked Running, so from inside a tenant’s code the platform being down looked like the function being slow. A run record needs a way to say whose failure an attempt was.
English Wikipedia lets software edit its pages on three conditions: its own account, a person named on that account’s page, and an approval for each task. The site can read all three. An agent pipeline’s audit line, actor and model ID, is readable only by the people who run the pipeline, and the site it wrote to is the reader with no copy.
A model vendor’s deprecation notice names the replacement and tells you to test against it before the retirement date, so every criterion written inside that window was written knowing the answer. Device regulators already settled what counts instead: criteria dated before the change.
A blameless postmortem is a trade: the engineer who broke it gives a full account of what they did and saw, and in return the account can’t be used against them. When an agent wrote the change, its account is already in the transcript. The person whose account isn’t is the one who approved the run.
Dry-run is the default in the operational layer I built around an unattended coding agent, and I’ve been reading more into that flag than it can deliver. A dry run is faithful up to the first side effect the agent reads the result of. After that it’s a run through a world the agent made up.
Google’s review tooling gives engineers a button that says the analyzer was wrong, and disables analyzers that collect too many of them. An agent never presses it, so a guard pipeline aimed at agents loses the only instrument that ever measured its own rules.
Run git blame on a line an autonomous agent wrote and it points at a service account. The commit is spotless. The one thing debugging actually needs, a person who remembers why, already exited.
Months ago I listed what a system needs before an agent can run unattended and treated the list as finished. It wasn’t. A stop button handles the runaway. It does nothing for the run that dies on step nine of fourteen.
I built audit logs and self-voiding approval gates to make an autonomous agent safe to run unattended, then realized I’d rebuilt a regulation the FDA shipped in 1997. The regulated world solved agent trust before agents existed.
Birgitta Böckeler’s harness-engineering framework treats iteration as human work. Someone watches, notices, edits the rules. That assumption breaks the moment agents run overnight. The next move is sensors with memory.