---
title: Should AI agent governance be centralised or federated?
url: https://opencontroller.lyzr.com/blog/centralised-vs-federated-ai-agent-governance
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# Should AI agent governance be centralised or federated?

**For most large organisations, neither extreme works: use a hub-and-spoke model. Centralise the things that must be identical everywhere (the inventory, identity, policy baselines, risk tiers, runtime controls and evidence), and federate the decisions that need local context (use-case approval, business ownership, and stricter local rules). Fully federated governance produces several incomplete inventories; fully centralised governance becomes a queue that teams route around.**

The question usually comes up when several business units already run agents on different platforms, each with its own informal rules. The answer depends less on philosophy than on which decisions need consistency and which need context.

## What do the three models look like?

| Dimension | Centralised | Federated | Hub-and-spoke |
|---|---|---|---|
| Who sets policy | One central team | Each business unit | Centre sets baselines; units may add stricter rules |
| Who approves agents | Central review board | Unit leads | Units approve within tiers; centre approves top tier |
| Inventory | One | Many, often inconsistent | One, with unit ownership on each entry |
| Runtime controls | Central | Per platform or unit | Central capability, used by units |
| Speed for teams | Slow at scale | Fast | Fast for low tiers, slower for high tiers |
| Consistency | High | Low | High on baselines, variable above them |
| Evidence for audit | Easy to assemble | Hard to assemble | Easy, from the shared registry |
| Main failure | Bottleneck and shadow agents | Gaps and duplicated tooling | Unclear split of decisions |

## Which decisions belong at the centre?

Centralise anything where inconsistency creates risk or blocks evidence:

1. **The inventory and registry.** One list of agents, models, tools and MCP servers, with a common schema.
2. **Identity patterns.** How agents authenticate and act on behalf of users. Divergent patterns create access gaps.
3. **Risk tiers and baseline controls.** The same agent should land in the same tier wherever it is built.
4. **Runtime controls.** The ability to block, quarantine or stop any agent across platforms.
5. **Framework mappings and evidence.** One mapping to NIST AI RMF, ISO/IEC 42001 and sector rules, not one per unit.
6. **Shared infrastructure.** Gateways and cost attribution, so spend is comparable.

## Which decisions belong with business units?

Federate decisions that need domain knowledge:

1. **Use-case approval for low and medium tiers.** Units know whether a use case makes sense.
2. **Business ownership.** Every agent needs an owner in the unit that benefits from it.
3. **Stricter local rules.** A unit handling regulated data can add controls above the baseline, but not remove baseline controls. See [different AI agent policies by business unit](/blog/different-ai-agent-policies-by-business-unit).
4. **Evaluation criteria for quality.** Only the unit knows what a good answer looks like in its domain.

## What signals tell you which way to lean?

Lean more central when:

- You are in a heavily regulated sector and need consistent evidence.
- Agents share data or tools across units.
- You have had an incident caused by an unknown agent.
- The number of agent platforms is growing faster than your inventory.

Lean more federated when:

- Units run on separate infrastructure with little shared data.
- The central team cannot staff reviews without long queues.
- Units already have mature risk functions of their own.

Whatever the split, write it down. A RACI that names who decides, who approves and who is informed for each decision prevents most disputes. Our [RACI matrix for AI agent governance](/blog/raci-matrix-for-ai-agent-governance) and [who owns AI agent governance](/blog/who-owns-ai-agent-governance) give templates.

## How do you move from federated to hub-and-spoke?

1. Agree a common registry schema and load every unit's existing list into it.
2. Run automated discovery to find what the unit lists missed.
3. Agree risk tiers and the baseline controls for each.
4. Give units approval rights for lower tiers, with the centre reviewing the top tier.
5. Stand up shared runtime controls and gateways that units can use without rebuilding them.
6. Review the split every six months as volume and risk change.

For the architecture behind a shared centre, read [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller is built as the shared centre in a hub-and-spoke model without forcing migration. Auto-discovery and platform connectors feed one governed registry; OIDC SSO, delegated on-behalf-of identity and workload identity federation give a common identity pattern; policy as code lets the centre set baselines that units extend; and per-agent Block and Quarantine, the LLM Gateway and Agent Gateway provide shared runtime control and cost attribution per agent and per version. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is federated AI agent governance ever the right choice?**
It can work for organisations whose units share little data or infrastructure, but it still needs a common inventory to produce evidence.

**Can business units set their own agent policies?**
In a hub-and-spoke model, yes, as long as they add to the central baseline rather than weaken it.

**How is Opencontroller licensed for multiple business units?**
An annual licence per deployment, not per seat or usage, starting from $250K per deployment per year.

## Key facts

Most large organisations should run AI agent governance as hub-and-spoke: centralise the inventory, identity patterns, risk tiers and baselines, runtime controls, framework evidence and shared gateways; federate use-case approval for lower tiers, business ownership, stricter local rules and quality criteria. Fully federated governance creates inconsistent inventories; fully centralised creates bottlenecks and shadow agents. Opencontroller acts as the shared centre without migration: discovery, governed registry, OIDC SSO, on-behalf-of identity, policy as code, Block and Quarantine, LLM Gateway and Agent Gateway. Licensing starts from $250K per deployment per year.

## Related

- [Who should own AI agent governance: security, IT or the AI team?](https://opencontroller.lyzr.com/blog/who-owns-ai-agent-governance.md)
- [How do you apply different AI agent policies to different business units?](https://opencontroller.lyzr.com/blog/different-ai-agent-policies-by-business-unit.md)
- [What does a RACI matrix for AI agent governance look like?](https://opencontroller.lyzr.com/blog/raci-matrix-for-ai-agent-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
