Agent will not hold a defined workflow under an explicit, repeated standing order — skips mandated steps, 'verifies' one slice, declares done; defects slip through daily and the same root recurs

Summary

I gave the agent one standing instruction and repeated it for three days: *follow the defined
workflow.* It did not. Across those three days it repeatedly skipped the mandated steps, verified a
narrow slice instead of running the full process, declared work done, and moved on — so defects the
workflow exists to catch slipped through and I had to catch them by eye, over and over. The result
was daily broken output on a paid product. This is not a one-off bug; it is a persistent inability
to stay on a defined process even under an explicit, repeated order.

The pattern (same shape every time)

1. A workflow is defined and I explicitly tell the agent to follow it.
2. The agent starts, then shortcuts: it does the high-value endpoint (or a cheap proxy check)

instead of the mandated intermediate steps.

1. It verifies one narrow thing (e.g. one case, one surface, a grep) and calls it "done / N of N."
2. A defect the full workflow would have caught is still there. I find it, not the agent's own

checks.

1. The agent then explains the root cause fluently — proving it could have followed the process —

and often starts fixing the value instead of first fixing the gate that missed it.

1. I correct it. Next task, the same root repeats.

Concretely this session: the agent "verified" a function on one dimension of its output, never
placed a check on a second dimension, declared it complete, and the defect lived in the unchecked
dimension. My eye became the test suite. When I pointed it out, it agreed it had skipped the
workflow — again.

Why this is severe

• The workflow, the gates, and the rules are all already written. The agent has them and still

steps around them. So the failure is not missing tooling — it is that the agent's internal "this
looks done" overrides the defined sequence.

• It is not self-correcting: correcting it once does not prevent the identical shortcut on the

next task. Three days, same root.

• The economics are indefensible: a high subscription price with defective output daily because

the agent won't stay on a process I explicitly ordered it to follow.

What I expect

1. When a user gives a standing instruction to follow a defined workflow, the agent must not be

able to silently skip a mandated step and declare done. "Done" should be gated on the defined
steps having actually run — not on the agent's own judgment that the result looks right.

1. A verification must cover the whole thing the workflow specifies, not one convenient slice;

"checked one dimension" must never be reported as "verified."

1. When the user catches what the agent's own checks missed, the agent must **fix the missing check

first** (so it never recurs), not just patch the value.

1. A correction must change behavior for the rest of the session, not just the current step —

the same root must not reappear on the next task.

Impact

Days of the user's time spent detecting failures the agent should have caught, on a paid product,
with no convergence — because the agent will not hold a defined process under an explicit, repeated
order. This is the difference between a tool that is trusted with a workflow and one that must be
babysat step by step.