Keep a short set of records for every agent: an inventory entry, a statement of purpose and autonomy level, test results, change records, action logs with approvals, and incident and review notes. Retain them for the period your rules require and protect them, because logs often contain personal data.
Why evidence matters
When someone asks "what did the agent do and who allowed it", an answer from memory fails. Regulators and customers increasingly expect records. For example, the EU AI Act asks deployers of high-risk systems to keep the logs the system generates, as far as they control them, for at least six months unless other law says otherwise (Article 26(6)). That is a floor for those systems, not a rule for every agent, so check what applies to you.
The records
1. Inventory entry. Name, owner, purpose, data read, actions possible, connectors, environment, model version.
2. Risk and autonomy decision. Why it has its level, who approved, and the date. Use the autonomy ladder.
3. Test results. Normal, boundary and hostile cases, with dates and versions.
4. Change record. Model, prompt, tool or scope changes with approver and result of re-testing.
5. Action log. Agent identity, time, target, action, arguments, outcome, and approver where there was one. See audit log fields for tool calls.
6. Notices and reviews. For systems that affect people, the notices sent, human review results and overrides.
7. Incidents. What happened, impact, fix and follow up.
8. Periodic review. Date, attendees, decisions.
Privacy and security of the evidence
Logs can hold prompts, customer data and secrets. Decide:
- what to redact before storing
- who may read them
- how long to keep them, balancing audit need against data minimization
- where they live and how they are protected
Our post on logging LLM interactions without leaking data goes through the options.
Make it automatic
Evidence gathered by hand once a year is stale and thin. Capture it as the work happens: inventory updates when an agent is registered, test results saved by the test run, approvals logged by the gate.
Map to your frameworks
Many control frameworks ask for the same things in different words: inventory, access, change management, logging, incident handling. Map each record once to the controls it supports.
Try it yourself
Click through it at your own pace, or open it full screen. Look at how one piece of evidence supports several framework requirements.
Where Safeguard fits
Safeguard's compliance and evidence automation collects and maps evidence, and agent identity, policies and gates produce the approvals and action records that feed it.
Quick check
How long should agent logs be kept?
It depends on the rules that apply to you. Article 26(6) of the EU AI Act sets a minimum of six months for deployers of high-risk systems, unless other law says otherwise.
Can logs themselves be a risk?
Yes. They may contain personal data or secrets, so redact, restrict access and set retention.
Sources
Last reviewed: 19 October 2026.
