# Agent Governance

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

The control layer that decides what AI agents may access, which actions need human approval, and how their work is logged, audited and reversed inside a business.

---
## Definition

Agent governance is the discipline of controlling what AI agents do inside a business: which data they can see, which tools they can call, what lands in memory, which actions require a human sign-off, and who can stop them when they go wrong. In 5ac's vocabulary it is the control plane: the central management layer between people, workflows and models that lets many agents operate safely as one business system. The target is not running the most agents. It is running auditable agents without chaining the whole workflow to a single vendor.

## Mechanism

The 5ac control-plane piece names four recurring failures when companies multiply agents ahead of governance: over-broad permissions (a sales agent seeing finance data), errors nobody can trace, memory scope leaking between customers, and no emergency brake. Six components answer them: identity and permissions; memory split by customer, department and workflow; model routing by cost and sensitivity; approval workflows; an audit trail capturing prompt, tool call, input, output and approver; and observability with rollback.

The second mechanism is process. The Kanban board turns each job into a card with an owner, completion criteria and a mandatory review checkpoint — review stops being improvised, and cycle time becomes measurable. Actions split into three tiers: agents work alone on low-risk tasks, agents propose and humans approve anything touching money or customers, agents only prepare data for the person who decides. The scholarly anchors: *AgentOps: Enabling Observability of LLM Agents* (arXiv:2411.05285, Dong, Lu and Zhu, submitted 8 Nov 2024) argues from a DevOps angle that autonomous, non-deterministic agents need their own observability pipeline; *A Survey on AgentOps* (arXiv:2508.02121, Zexin Wang et al., 4 Aug 2025) organizes the field's categories, challenges and directions.

## Application context

Compliance pressure is named in 5ac's own coverage: the EU AI Act reached an important enforcement milestone on 2 August 2026, and OWASP Agentic 2.0 now treats agent risk as its own security category. Seven product requirements screen any agent platform before scaling: workflow-based routing rather than vendor routing; clear permission boundaries; approval gates for sensitive actions; audit trails operations staff can actually read; cost ceilings and cost attribution per workflow; model portability; and sovereignty at the deployment layer. One demo question cuts through the pitch: "Six months from now, if I swap models, tighten access and keep my workflows, how long does that take me?"

## Sources fetched

- [AI agent control plane and governance](/en/blog/ai-agent-control-plane-governance/) — the four risks and six components — accessed 2026-10-19
- [Hermes Kanban for business automation](/en/blog/business-automation-hermes-kanban-vietnam/) — cards, owners, mandatory review checkpoints — accessed 2026-10-19
- [Agent Control Plane for Vietnamese SMBs — 7 requirements](/en/blog/agent-control-plane-vietnam-smb/) — the anti-lock-in requirement list — accessed 2026-10-19
- [Sovereign AI agents for Vietnamese SMBs](/en/blog/sovereign-ai-agents-vietnam-governance-compliance/) — EU AI Act milestone, OWASP Agentic 2.0, three action tiers — accessed 2026-10-19
- <https://arxiv.org/abs/2411.05285> and <https://arxiv.org/abs/2508.02121> — the AgentOps papers — verified via arXiv API 2026-10-19

## Related

- [What is an AI Agent?](/en/wiki/ai-agent/)
- [What is a sub-agent?](/en/wiki/sub-agent/)
- [Agent memory](/en/wiki/agent-memory/)
- [Model Context Protocol (MCP)](/en/wiki/mcp/)
- [Benchmark suites for AI agents](/en/wiki/benchmark-suite/)