# Audit log

> VI: https://5ac.vn/wiki/audit-log/ · EN: https://5ac.vn/en/wiki/audit-log/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C13

The time-ordered record of every action that matters in a system or AI agent stack: who triggered it, when, what data was touched, the result and its cost.

---
## Definition

An audit log is a time-ordered record of every action a system or an agent takes that matters: who triggered it, when, which tool was called, which data was touched, what came back and what it cost. When those records chain together and let you walk a result back to its cause, the chain is the audit trail. In an AI agent stack the logged actor is no longer a person: an agent that reads an inbox, edits a CRM record or sends a quote leaves the same kind of trace as an employee. The audit log is one component of an agent orchestration layer — the permission and sandbox boundaries belong to the agent-orchestration entry; this entry is about the record itself.

## Mechanism

The test for a usable record is simple: someone else can reconstruct the incident without asking the agent. The 5ac orchestration piece lists three things a log buys you — without one you cannot investigate incidents, improve ROI, or build internal trust in agent workflows. The platform guide adds the enforcement layer: per-agent credentials, scoped API access, and an audit log on every external call. Two operating rules come with it: any workflow touching customer data needs a traceable log plus a defined memory-retention period, and logs, dashboards and a kill switch go in before volume goes up. The buying test collapses to four checks: policy, logs, approvals, rollback.

| When it pays | The question it answers |
| --- | --- |
| Incident | Which step went wrong, on what input? |
| Cost | Which workflow or agent is spending tokens? |
| Compliance | Who approved that write to customer data, and when? |

## Application context

Audit trail turns build-versus-buy into a concrete question: build your own when every step must produce a trail in a specific government format — not merely because "we process invoices". Platforms that already ship audit trails, explainability and human override hold the advantage from day one, the platform guide notes. For Vietnamese SMBs the ROI line item is traceable errors: an error you can trace is an error you fix at the root instead of patching at the symptom.

## Sources fetched

- [AI Agents Platform: Complete Guide for Business (2026)](/en/blog/ai-agents-platform/) — per-agent credentials, audit logs on every external call, government-format trails — accessed 2026-10-19
- [Secure Agent Orchestration for Vietnamese SMBs: Identity, Audit Logs, and Control](/en/blog/secure-agent-orchestration-layer-vietnamese-smbs/) — incidents, ROI and internal trust as the three jobs of a log — accessed 2026-10-19
- [What Is an AI Agent Control Plane? Governance for Safe Enterprise Automation](/en/blog/ai-agent-control-plane-governance/) — traceable logs, retention limits, kill switch before scale — accessed 2026-10-19

## Related

- [Agent orchestration](/en/wiki/agent-orchestration/)
- [Agent Governance](/en/wiki/agent-governance/)
- [AI Agent](/en/wiki/ai-agent/)
- [Token cost control](/en/wiki/ai-token-cost/)