Contextual Agentic

Security Logs Need a Decision Record

Tool-call logs say what an agent attempted. An assurance record also explains authority, policy, approval, system outcome and independent verification.

·5 min read

Written as at 22 October 2026.

observabilityauditagent-securityassurancesecurity-operations

Most agent demos can show a transcript. Fewer can answer a harder question: why did the system permit this exact action, and what actually happened after the tool returned? A transcript is a useful clue, but it is not an audit record. It may omit the user identity, the policy version, the data that was withheld, the approval and the state of the system of record after the action.

I would build the record around the decision boundary rather than around the model’s prose. The model can propose. The gateway evaluates. A person may approve. The system of record commits or refuses. An independent observer records the outcome. Each step should have its own timestamp and identifier so an investigator can reconstruct the sequence without trusting one component’s story.

A request that looked successful

In an illustrative MAAH Group workflow, an agent proposes blocking a supplier account after a suspicious login. The tool returns HTTP 200. A dashboard labels the task “complete”. But the supplier system accepted the request as a queued change and later rejected it because the account had already been merged. The agent’s log is true about the tool response and wrong about the business outcome.

An assurance record separates four claims: the request was made, the gateway allowed it, the destination accepted it, and the intended state was observed. A reliable dashboard reports which claims have evidence and which remain pending.

Record field Why it matters
Human, agent, task and tenant identifiers Prevents a shared service identity from hiding who initiated the action.
Source IDs and classifications Shows what context was used without copying raw sensitive material into every log.
Typed command, policy and approval identifiers Binds permission to one specific operation rather than a broad narrative.
Tool response and system-of-record state Separates acceptance from actual commitment.
Verification result and retention policy Supports later audit, replay and incident investigation.

Do not log passwords, bearer tokens, raw private keys or unredacted personal records to make the trail “complete”. Record a reference, keyed digest or access-controlled evidence object where appropriate, then test the redaction with a known dummy secret. A diagnostic that leaks credentials is itself a security incident waiting to happen.

What a real-time dashboard should show

The useful view is a timeline of proposed → checked → approved → submitted → committed → verified, with refused and timed-out branches visible. It should let an analyst pivot from an anomalous tool call to the underlying identity, policy and evidence, and it should highlight missing steps instead of inferring success from silence.

Before automating response, inject a tool timeout, a conflicting approval and a destination rejection. Confirm that the dashboard shows an unresolved or failed state, and that an agent cannot rewrite the record after the fact. This is how logging becomes assurance rather than decoration.

SAE’s evidence and assurance chapters build the full control chain. See the books page for the edition’s status.

References

  1. The Contextual Secure Agentic Enterprise, first edition, Chapters 18 and 20B (prepublication manuscript, 22 October 2026).

← All articles

Related