Enough to reconstruct any action the agent took: not “what did it do” as a summary, but “show me exactly what it did, when, with what inputs, who authorised it, and what the outputs were used for.” MING Labs’ operating standard answers that with ten fields, traced for every agent action, with the whole fleet logged: 100 percent of tasks carry a timestamp, a cost, and an outcome, and no work happens off-ledger.[S1]
01The ten fields
Per action, the trail records:
- Agent identity and version of the agent that acted
- Triggering event: who or what initiated the action
- Job-description scope in force at the time of execution
- Authority level, with the timestamp of its last change
- Tool calls, with their parameters
- Data accessed: sensitivity tier, source, and any redaction applied
- Output: the artefact produced and its destination
- Cost and duration of the run
- Human review: who reviewed, when, and what they decided
- Downstream consumption chain: what the output was subsequently used for
The organising principle matters more than the count. Every entry maps to a named element of the agent’s job description: the scope field points at the mandate, the authority field points at a logged grant, the review field points at the named owner. The audit trail is the job description in motion, not a stream of opaque events.[S1] That mapping is what separates an answer from a log dump, and it is what generic “log everything” platforms cannot produce, because they have no job description to map to.
02The question the trail exists to answer
The test case is a regulator’s request, because it is the hardest version of the question anyone will ask. A financial or healthcare supervisor does not ask whether you log; they ask you to show exactly what a specific agent did in a specific case, on whose authority, and where the output went. If that answer takes longer than 24 hours to produce, agents are not deployable in regulated work, full stop.[S1] The same ten fields answer the friendlier versions of the question that arrive first: a client asking how their data was used, an IT security team scoping an incident, or a named owner deciding whether an agent has earned more autonomy.
Regulation is converging on the same requirement. Article 12 of the EU AI Act requires high-risk AI systems to technically allow automatic recording of events over their lifetime, applicable from 2 August 2026,[S2] and Article 26 obliges the deployer to keep those logs for at least six months.[S3] The direction of travel is unambiguous: traceability is becoming a property a system must have, not a report a team assembles afterwards.
03Retention and review, as operated
A trail nobody reads is a liability with good intentions, so the standard includes cadence and retention, not just fields. At MING Labs, operational logs are kept 90 days, security events 12 months, and client-work records for the contract term plus two years. An automated audit sweeps every agent weekly for drift, security anomalies, and performance regressions; high-severity work gets an additional monthly manual review.[S1] Field nine closes the loop: the review is itself recorded in the trail it reviews.
04The claim, scoped
Scope
- The standard: ten fields per agent action, each mapped to a named element of the agent’s job description; 100 percent of tasks logged with timestamp, cost, and outcome.
- Where it runs: MING Labs’ own production fleet, since January 2026.
- Retention as operated: 90 days operational, 12 months security, contract plus two years for client work; review weekly automated, monthly manual for high-severity work.
- Not claimed: a compliance certification, or that this satisfies any specific regulator’s requirements. Article 12 and Article 26 of the EU AI Act are cited as the regulatory direction, not as an attestation of conformity.
- Disclosure: structure only. Log contents, client identifiers, and internal system names are not published.
The trail is also what makes the rest of the operating model honest. Accountability assignments are checkable because authority changes are logged events, as described in who is accountable when an AI agent makes a mistake . Retirement decisions rest on recorded outcomes rather than impressions, which is how MING Labs knew its first agent had produced 764 messages and zero acted-on output before shutting it down . And the security half of the story, why the tool calls in field five are all mediated by a credential-holding gateway, is covered in should AI agents hold their own credentials .
