Yes. A compliance agent that shares a brief, sources or owner with the agents it checks will share their blind spots. Give it a separate brief and owner, read-only access to the other agents' records, no power to change their bounds, and its own certificate. A compliance officer still decides what counts as a breach.
Why does separation matter so much?
Segregation of duties is old advice for a new kind of worker. The person who prepares a payment should not approve it; the team that sells a product should not be the only one checking how it was sold. The same logic applies when the worker is an agent.
There is also a reason specific to agents. If the checking agent reads the same policy extract, uses the same sources and was configured by the same team, its errors are correlated with the errors it is meant to catch. A misread rule produces a wrong action and a clean check on the same day. Separation is how you make failures independent, which is the whole point of a second line, as the guide to AI compliance agents sets out.
What does separate mean in practice?
- A separate brief and a separate owner, sitting in compliance rather than in the team that runs the first-line agent.
- Read-only access to the checked agent's filed verdicts, drafts and the sources each one used.
- No write access to the checked agent's bounds, brief or outputs. It reports; it does not fix.
- Its own copy of the rules, taken from the policy library, not from the first-line agent's brief.
- Its own certification, tested on cases where the first-line agent was known to be wrong.
A first-line agent can still check its own work before it acts, and it should. That is a quality step, not a compliance check, and it should not be reported as one.
What may the compliance agent prepare, and what stays with a person?
| Task | The compliance agent | The compliance officer |
|---|---|---|
| Sample the first-line agent's actions | Selects and reviews the sample | Sets the sampling rule |
| Compare actions against policy | Flags each exception, citing the rule and the record | Decides whether it is a breach |
| Spot patterns across many actions | Drafts a summary with examples | Decides whether to restrict the agent |
| Recommend a change to bounds | Drafts the recommendation | Approves it; the first-line owner applies it |
| Report to the board or regulator | Drafts the figures and narrative | Signs and submits |
What does this look like in an onboarding team?
Take a business bank where a first-line agent prepares onboarding files and requests identity checks on customers at depth 3, within limits. A separate compliance agent, owned by the second line, reviews a daily sample of those files.
One morning it flags that six files for companies with overseas directors went through without the extra check the policy requires. It cites the policy clause, links each file and notes that the first-line agent's brief carried an older version of the rule. The compliance officer confirms the breach, the onboarding lead fixes the brief, and the first-line agent is held at depth 2 until it is recertified. Had the same agent checked itself against the same outdated brief, all six files would have passed.
Does the same apply to other checks?
The principle holds wherever a check exists to catch a different party's mistake. It is plainest in conflict of interest checks and anti-bribery checks on third parties, where the checker must be independent of the people and processes being checked.
It applies with less force to operational checks run within a team, such as end-of-day operations checks. There, the checking agent can sit with operations, provided it is not the same agent that did the work.
Where Pryme fits
In Pryme each agent has its own role scope and its own certificate, and every verdict is filed in an append-only record. A compliance agent can be given read access to another agent's record without any ability to change that agent's bounds; only a named person can do that.
Questions people ask
Can the compliance agent stop the first-line agent directly?
It is better that it cannot. It should flag and recommend; a named person restricts the agent, so there is clear accountability for stopping work.
Should the two agents use different models?
Where practical, it helps, because it makes shared errors less likely. Separate sources, rules and owners matter more than the model.
Who certifies the compliance agent?
The compliance function, tested on past cases where the first-line work was known to be wrong. A checker that has never been shown a mistake has not been tested.