Article contents
Regulatory information can change. This resource is educational and should not be treated as legal, clinical, or regulatory advice. Consult the applicable primary guidance and qualified professionals for decisions affecting your organization. Last reviewed November 18, 2026.
Traces are not an audit trail
Most teams running agents already have tracing: every prompt, completion, tool call and latency figure, kept so engineers can debug. That is valuable, and it is not the same thing as an audit trail.
A trace is written for the engineer who built the agent. An audit trail is written for the people who have to account for it: a manager reviewing an incident, a customer asking why their refund was declined, an auditor, or a regulator. They do not want token counts. They want to know who decided, on what basis, and what happened next.
Debug trace versus audit trail
What the rules ask for1,2,3,4,5,6
The EU AI Act requires high-risk AI systems to allow automatic recording of events over their lifetime, so that risky situations can be identified and operation can be monitored. Providers must keep those logs, and deployers must keep the ones under their control, for at least six months unless other law requires otherwise. The high-risk rules apply from 2 December 2027.
Colorado's 2026 law gives people affected by an adverse automated decision the right to ask for more information about the technology and its inputs, and to request meaningful human review. A deployer cannot answer that without a record of what the system used and did.
NIST's work on agent identity points to the same need from the security side: actions must be traceable to a specific agent with specific, scoped authorization, not to a shared account. Singapore's agentic AI framework adds a step many teams skip: audit whether your human approvals are still effective.
Record this for every action an agent takes
Each entry should carry a timestamp from a single trusted clock.
- The agent's identity and the version of its instructions and permissions at that moment
- The person or team that owns the agent
- The run, and what started it: a schedule, a message, a person
- The permission that allowed the tool call
- The tool and the exact parameters, with sensitive values redacted
- References to the sources the agent relied on, rather than full copies
- The result returned by the outside system
- Whether approval was required, and which rule required it
- The reviewer, their decision, any edits, the time and their reason
- Whether the action can be undone, and the reference needed to undo it
Four properties of a record you can trust
Written outside the agent
The platform records what happened. An agent that misread a situation, or read a hidden malicious instruction, is the wrong party to write its own history.
Append-only
Entries are added, never edited. Corrections are new entries that point to the old ones, so the record shows its own history.
Attributable
Every entry names an agent identity or a person. Shared service accounts make it impossible to say who acted.
Minimal
Store references and redacted values instead of copying sensitive data into logs. Separate the event from the payload so each can follow its own retention rule.
How long to keep it2,3
There is no single answer. Under the EU AI Act, six months is the minimum for logs of high-risk systems, and the Act defers to other law, including data protection law, where it says otherwise. Industry rules, contracts and your own incident history may call for longer. Data protection law may call for less when entries hold personal data.
A practical pattern: keep the skeleton of every event, who acted, what was done, when and with what outcome, for the longest period any obligation requires, and delete or redact sensitive payloads on a shorter clock. Decide per record type, write the decision down, and make sure the trail does not disappear when an agent is deleted.
Common mistakes
- Letting the agent write its own audit log
- Running agents under a shared human or service account
- Recording the prompt but not the action that was actually carried out
- Approvals that happen in chat and never reach the record
- No copy of the agent's permissions as they were at the time
- Logs deleted along with the agent
- Personal data copied into every log line
- No way to export the record when someone asks for it
Where Quaspar fits
Quaspar is the governance layer for AI agents, and its activity record is written by Quaspar, not by your agents. Every security-relevant event is recorded with who, what, when and the outcome: every agent run and tool call, approvals and rejections, permission changes, releases, cancellations, pauses and deletions, as well as sign-ins and exports. Sensitive details used during a run are removed from stored run history when the run finishes, on retention windows you can configure, and your organization's data can be exported at any time.
Quaspar does not decide which records a law requires you to keep, and it is not SOC 2 certified. It gives you a complete, attributable record to build on. Quaspar is in private preview.
Next step
Quaspar is the governance layer for AI agents and is in private preview. Request access to set agent permissions, route important actions to people and keep a record of what every agent did.
Request Studio accessSources
Authoritative references
Rules, standards, and guidance can change. The links below were reviewed on November 18, 2026.
- 1European Commission, AI Act Service Desk. AI Act, Article 12: Record-keeping
- 2European Commission, AI Act Service Desk. AI Act, Article 19: Automatically generated logs
- 3European Commission, AI Act Service Desk. AI Act, Article 26: Obligations of deployers of high-risk AI systems
- 4Colorado Legislative Council Staff. 2026 Summary of Legislation: Information Technology and Artificial Intelligence
- 5NIST National Cybersecurity Center of Excellence. Accelerating the Adoption of Software and AI Agent Identity and Authorization (concept paper)
- 6Infocomm Media Development Authority, Singapore. Model AI Governance Framework for Agentic AI, version 1.5
