Agent Risk Register

Agent risk register

A risk register works only if it changes behaviour. Most decay into a spreadsheet reviewed annually, listing risks nobody owns. Agents can keep it current because the maintenance is repetitive and the review cadence is exactly what humans skip.

Team

• Scanner: proposes new risks from incidents, changes, and external signals.
• Assessor (risk-analysis, pre-mortem): rates likelihood and impact with a stated basis.
• Chaser: tracks mitigation owners and dates, escalating what has stalled.

Shape: continuous intake with a scheduled review of the whole register.

Method

1. Write risks as specific consequences. Losing the main cloud region for a day, not cloud risk. Vague entries cannot be assessed, mitigated, or closed.
2. Rate likelihood and impact with a stated basis. Whatever scale you use, the reasoning must be recorded, since unexplained ratings are unarguable and therefore unimprovable.
3. Give every risk an owner and a decision. Mitigate, accept, transfer, or avoid. An accepted risk is a legitimate outcome and must be recorded as a decision rather than an oversight.
4. Track mitigations as work with dates. A mitigation with no owner or date is an intention (see agent-compliance-desk).
5. Feed the register from real events. Incidents, near misses, audit findings, and dependency changes each propose entries automatically (see agent-postmortem-council).
6. Force stale entries back to a decision. Anything unreviewed past a threshold is escalated, since a register full of untouched entries is worse than a short honest one.
7. Review the top risks on a fixed cadence with a human. The review produces decisions, not acknowledgements.

Run it

In Claude Code, keep the register as a single tracked file, run the scanner on a schedule over incident and change records, the assessor on new entries, and the chaser as a sweep that escalates. A human runs the periodic review. Port to LangGraph with an intake node and a scheduled escalation node.

Signals it works

• Every open risk has an owner, a decision, and a review date.
• New risks arrive from real events rather than from brainstorming.
• Accepted risks are explicitly accepted, with a name attached.

Boundaries

Agents maintain and chase; humans own risk decisions, which are judgement calls with real consequences. Regulated risk frameworks have their own requirements and are not satisfied by this register. Some risks, particularly around people and legal exposure, must not be recorded in shared systems at all.