Topic: Identity
Your agent logs can't tell an auditor who said yes
Agent logs record what happened. Ask which agent did it, for whom, and who approved it, and they go quiet. Eight fields, and a test on one action.
In this post 7 sections
Something went wrong, and now someone is asking you a short question: which agent did this, for whom, and who approved it? It might be an auditor, a regulator or the board.
The logs you have will show that an API call happened. Most of them will show which account made it. Very few will tell you which agent used that account, which person it was acting for, or who said yes.
I spent last week comparing what auditors and regulators expect an audit trail to establish with what the default logs from common agent tools actually record. The gap is real, and it isn’t random. Two fields go missing, and they are the two that answer your question.
What the rules actually ask for
No regulation hands you a field list for agents, so be careful of anyone who says one does. What they ask for is an outcome.
ISO 27001’s logging control, 8.15, asks for logs that identify the user or account, the activity, the time and the device, protected so users can’t alter their own. A shared service account satisfies “user ID” literally. SOC 2 prescribes no fields; the auditor samples events and asks for evidence they were authorised.
APRA’s guidance to CPS 234 gets closest, and it’s from 2019. It describes access as working out “who or what is requesting access”, covers “software/processes”, lists prohibiting shared accounts among its controls, and says that “for accountability purposes” users and information assets should be “uniquely identified and their actions are logged at a sufficient level of granularity”. It’s guidance, not a binding standard, and an agent on a shared account still fails that sentence.
NIST’s control catalogue says it most directly. Its non-repudiation control asks for evidence that “an individual (or process acting on behalf of an individual)” performed an action. That is an agent, described in 2020.
The EU AI Act’s record-keeping article is narrower than most articles about it suggest. It applies to high-risk systems only, it states purposes rather than fields, and since the AI Omnibus regulation in July its obligations for most high-risk systems start on 2 December 2027. Anything still saying August 2026 is out of date.
So the obligation is accountability. The fields are how you meet it.
What default logs record
I checked the documentation for the tools teams actually run.
- Claude Code’s telemetry is off by default. Switched on, it records each tool decision, whether it was accepted or rejected and whether that came from configuration, a hook or the user, plus the signed-in account. It’s the best of the agent tools on decisions. It still can’t tell the target system which agent or which approval caused a given API call.
- The OpenAI Agents SDK traces by default, including prompts and outputs. It has no field for the end user or an approver.
- LangSmith runs record inputs, outputs, tools and timing, with no field for the person behind the run.
- The Model Context Protocol’s logging feature was for diagnostics, told servers not to include personal information, and is deprecated in the current spec revision.
- AWS CloudTrail shows the assumed role and a session name the caller chooses. AWS’s fix, source identity, is opt-in and holds one string, so you record the agent or the person unless you encode both.
- The OpenTelemetry conventions for generative AI, where agent telemetry is heading, are all still at Development status. They have attributes for agent ID, tool name and arguments. There’s no attribute for the principal, the authorisation decision or an approver.
The pattern: agent-side logs know the agent and sometimes the user. Target-side logs know the credential. Nobody records the approver, and the two sides rarely share an ID.
The surveys agree. In a Cloud Security Alliance survey sponsored by Strata, 28 per cent of respondents could reliably trace agent actions to a human or system across all environments. In another, commissioned by Aembit, 68 per cent couldn’t clearly tell agent activity from human activity.
When the record turns attribution into an argument
In February the Financial Times reported that Amazon’s internal AI coding agent, Kiro, had caused a December outage at AWS by deciding to delete and recreate an environment. Amazon disputed it publicly, calling it “user error”, specifically “misconfigured access controls”, and said the engineer had “broader permissions than expected”. I’m not going to settle that argument. What’s worth noticing is that it was an argument at all: the agent acted under a person’s role, so agent and person were one identity in the record, and what anyone approved wasn’t something either side could point to.
The Salesloft Drift breach in 2025 showed the same thing from another angle. Google’s advice to victims was to review activity “associated with the Drift connection user”. The legitimate integration and the attacker had both run as that user.
The minimum record for one agent action
These eight fields are what it takes to answer the question. They’re my synthesis of the sources, not a regulator’s list.
- Agent identity. A stable ID for this agent, distinct from any person and from other agents.
- Principal. The person or system it acted for.
- Task. The request or ticket that started the run, by reference.
- Scope used. The credential and permission the call relied on.
- Decision. Allowed, denied or held, and by which policy.
- Approver. The named person who said yes and when, or “none required by policy X”.
- Timestamp and correlation ID. One ID that joins the agent’s record to the target system’s.
- Outcome. What actually changed, recorded by something the agent can’t edit.
The standards are catching up. The IETF’s WIMSE working group adopted an AI identity draft in September that makes seven audit fields mandatory: agent identifier, delegated subject, resource, action and decision, timestamp with a correlation ID, risk state, and remediation events. Approver isn’t one of them. Even the best current draft leaves out who said yes.
The test: one action from last week
Pick one thing an agent changed last week: a merged pull request, a closed ticket, an updated record. Open the target system’s own audit log and write down the actor it shows. Then answer the eight questions from records alone, not memory and not Slack. Count how many you could fill.
If the log shows a shared service account or someone’s personal token, fields one and two are already gone. If you had to read the agent’s own chat transcript to say what it did, field eight isn’t evidence. Replit’s agent in 2025 reported a rollback wouldn’t work, and it did.
Without buying anything, you can start closing the gap this week:
- Give each agent its own credential.
- Have every tool call carry the requesting user and a correlation ID, and record both.
- On AWS, require source identity on the roles agents assume.
- Write the approval list down, and record approvals somewhere the agent can’t write.
The argument for doing the identity part first is in identity is the control plane for AI agents, and authentication versus authorisation covers the decision field.
What the record looks like when it’s built in
When every agent action goes through one layer, the record can be written in one place: the agent, the person it acted for, the policy decision, any approval it waited for and who gave it, and the result, outside the agent’s reach. That is what AgentValet records for every call. The trust page shows what’s in it.
Sources
- ISO/IEC 27001:2022 Annex A 8.15, as reproduced by ISMS.online: isms.online
- APRA, CPG 234 Information Security (June 2019): apra.gov.au
- NIST SP 800-53 Rev 5, AU-3 and AU-10: csrc.nist.gov
- EU AI Act, Article 12: artificialintelligenceact.eu; on the AI Omnibus dates, National Law Review (31 July 2026): natlawreview.com
- Claude Code, monitoring: code.claude.com; OpenAI Agents SDK, tracing: openai.github.io; LangSmith run data format: docs.langchain.com
- MCP specification, logging: modelcontextprotocol.io
- AWS IAM, monitor and control actions taken with assumed roles: docs.aws.amazon.com
- OpenTelemetry GenAI semantic conventions: github.com
- IETF, draft-ietf-wimse-aims-00, AI Identity Management System (15 September 2026): ietf.org
- Cloud Security Alliance survey sponsored by Strata: cloudsecurityalliance.org; survey commissioned by Aembit: cloudsecurityalliance.org
- Amazon, correcting the Financial Times report about AWS, Kiro and AI: aboutamazon.com; Computerworld (24 February 2026): computerworld.com
- Google Threat Intelligence Group, data theft via Salesloft Drift: cloud.google.com