FIELD NOTE / HUMAN-IN-THE-LOOP OPERATIONS
Five agent failures that deserve human escalation
Automation should continue when the next action is reversible and well-supported. It should stop when evidence, authority, or recovery becomes uncertain.
Source conflict
Two credible sources disagree on a fact that changes the action. Identify source dates, conflict, and decision risk.
Tool success without task success
An API returns 200, but the intended business outcome did not occur. Verify the state change, not merely the response code.
Authorization ambiguity
The agent can technically execute an action but lacks clear purchaser or operator authority. Stop and obtain accountable approval.
Silent scope expansion
A bounded task begins to require extra systems, sensitive data, or irreversible actions. Re-scope before proceeding.
Recovery-loop repetition
The agent retries variants of the same failed step without new evidence. Preserve logs, stop repetition, and route an exception record.
MUJ428's $19 Evidence Check handles one disputed claim. The $49 AI-Agent Exception Triage handles one blocked workflow.