Execuria · Oversight
Agent oversight
Oversight is the discipline that keeps agent contributions accountable to a corporate team's standards. It is not surveillance for its own sake, and it is not a single approval button. It is a set of supervision protocols that are light enough to be followed every day and strong enough to answer a hard question months later.
The four supervision protocols
Boundary review
At each hand-off, a named human inspects the work for the specific failure modes agreed in advance, and records the inspection in one line.
Sampling
High-volume work is reviewed by sample rather than in full. The sample size and the trigger for widening it are agreed in writing, not guessed.
Escalation
Any doubt is raised to a role empowered to pause the work. Escalation is treated as a healthy signal, never as a failure of the agent or the person.
Periodic deep read
On a slower cadence, a senior reviewer reads a larger body of output end to end, looking for drift that sampling cannot see.
We insist that oversight produce a trace. The trace need not be elaborate: a dated note, a short tag, a register entry. Its value is not compliance theatre; it is memory. When the business changes, the trace lets a new owner reconstruct what was checked and why.
We are equally firm that oversight must not become a choke point. A protocol that requires three approvals for low-consequence work will be abandoned within a month, and an abandoned protocol is worse than none, because everyone believes it is operating.
Oversight that is too heavy is quietly dropped. Oversight that leaves no trace did not happen.
Proportionality in practice
Not every contribution deserves the same scrutiny. We classify work by consequence — low, moderate and material — and attach protocols accordingly. Low-consequence chorus work draws sampling. Moderate work draws boundary review. Material work — anything touching money, safety, law, or a customer's rights — draws boundary review plus a named senior sign-off plus a periodic deep read.
| Consequence | Typical signals | Minimum supervision |
|---|---|---|
| Low | Internal, easily reversed, no external effect | Sampling review with an agreed sample |
| Moderate | Visible to colleagues, moderately costly to reverse | Boundary review with a recorded trace |
| Material | Money, safety, legal exposure, or customer rights | Boundary review, senior sign-off, deep read |
Who supervises the supervisors
A supervision protocol creates its own risk: the reviewer may become a rubber stamp. We guard against this with rotation, with a second reader on material work, and with a simple habit of keeping a small register of escalations. A register that is always empty does not prove the ensemble is flawless; it usually proves the escalation path is not trusted.
- Agree consequence bands. Classify the unit's work before writing any protocol.
- Name reviewers. Every boundary gets a person, not a team.
- Set the trace. Decide what a completed review looks like, in one line.
- Define escalation. Name who may pause work, and make pausing blameless.
- Read the register. Review escalations and near-misses on the model's cadence.
Oversight is where a collaboration model proves itself. It is also where most organisations are thinnest. Our diagnostic deliberately spends its time here, and our measurement notes describe how we watch oversight quality over time.
For the cast these protocols apply to, see the role atlas. For how oversight is scheduled and reported, see measurement notes.