---
title: Should production and test AI agents be governed in one control plane?
url: https://opencontroller.lyzr.com/blog/govern-production-and-test-ai-agents-in-one-control-plane
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# Should production and test AI agents be governed in one control plane?

**Yes. Put development, test and production agents in one control plane, tag every agent, key, tool and MCP server with its environment, and let policies tighten as an agent moves towards production. Separate control planes per environment hide the agents most likely to leak: the test ones running with real credentials.**

Teams often start governance with production only. The reasoning sounds fine: production is where customers are. In practice, test and UAT agents are where the surprises live. They are built fast, share API keys, call real tools "just for a demo", and are rarely switched off. This post sets out a tagging scheme and the policy differences that make one control plane work across environments.

## Why not run a separate control plane per environment?

Three problems show up when each environment has its own governance stack:

- **Agents fall between them.** A developer's prototype that quietly gets real users is in neither "test" nor "production" governance.
- **History breaks at promotion.** If an agent is re-registered in production, its evaluation results, owner and change history stay behind in the test system.
- **Evidence is fragmented.** Auditors ask for one inventory. Three inventories mean three exports and a reconciliation spreadsheet.

One control plane with environment as an attribute fixes all three. Promotion becomes a version change on the same registered agent, so the owner, evaluation scores and change log carry forward.

## What should the environment-tagging scheme look like?

Keep it small and mandatory. Every governed object gets the same tags:

| Tag | Allowed values | Applies to |
|---|---|---|
| `env` | `dev`, `test`, `uat`, `prod` | Agents, agent versions, model keys, tools, MCP servers |
| `owner` | Employee ID from your identity provider | Agents and tools |
| `data_class` | Your existing classification (for example public, internal, confidential, restricted) | Agents and tools |
| `cost_centre` | Finance code | Agents and model keys |
| `lifecycle` | `active`, `deprecated`, `retired` | Agents and versions |

If you already emit OpenTelemetry, reuse the `deployment.environment.name` resource attribute so traces and inventory share one value. Reject registration when `env` or `owner` is missing; untagged agents are the ones nobody can triage later. See [how to assign owners using your identity provider](/blog/assign-owners-to-ai-agents-with-your-identity-provider).

## Which policies should differ by environment?

The value of one control plane is that policies can differ while visibility stays the same. A typical starting point:

| Policy | Dev | Test / UAT | Production |
|---|---|---|---|
| Data allowed | Synthetic or public only | Masked samples | Classified data per approval |
| Write-capable tools and MCP servers | Sandbox endpoints only | Sandbox plus named test systems | Approved list only |
| Model access | Any approved model | Approved models | Approved models with fallback configured |
| Spend cap per agent | Low daily cap | Moderate cap | Budget with alerting by owner |
| Evaluation gate | None | Multi-turn and red-team evals run | Pass thresholds required to promote |
| Runtime guardrails | Logged | Logged and alerted | Enforced (block or redact) |
| Quarantine on deviation | Manual | Automatic alert | Automatic quarantine per policy |
| Change approval | Owner | Owner plus reviewer | Owner plus risk or platform sign-off |

Two rules prevent the most common incidents. First, **production credentials never appear in dev or test**; a discovered production key on a dev-tagged agent should raise a posture finding. Second, **a dev agent with external users is mis-tagged**; usage from outside the build team should trigger a review of its tag.

## How does promotion work in practice?

1. The owner registers the agent once, tagged `dev`.
2. Each change creates a new version in the registry.
3. Moving to `test` runs the evaluation suite against that version.
4. Moving to `prod` checks pass thresholds and the approval rule, then switches the tag on that version.
5. The previous production version stays registered so you can roll back.

Because cost is attributed per agent and per version, you can also compare the cost of the candidate version with the current one before promotion. See [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents).

## How Opencontroller does this

Opencontroller keeps one governed registry of agents, versions, tools and MCP servers, with policy as code so environment-specific rules live alongside each other and can be reviewed like any other code. Multi-turn and red-team evals and LLM judges give the evaluation gate; the LLM Gateway attributes cost per agent and per version; rogue-agent handling quarantines or kills agents that deviate. It is licensed per deployment, not per seat or usage. Read more in [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane) and [what is Opencontroller](/blog/what-is-opencontroller), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should test AI agents be in the same governance inventory as production agents?**
Yes. Test agents often hold real credentials and tool access. One inventory with an environment tag keeps them visible while allowing lighter policies.

**Which AI agent policies should differ between dev and production?**
Data allowed, write-capable tools, spend caps, evaluation gates, guardrail enforcement and quarantine thresholds.

**What happens to an agent's history when it moves to production?**
If promotion is a version change on the same registered agent, its owner, evaluation results and change log carry forward.

## Key facts

Recommended pattern: govern development, test/UAT and production AI agents in one control plane, with a mandatory environment tag on every agent, model key, tool and MCP server, and policies that tighten by environment. Promotion between environments should be a version change on the same registered agent, so evaluation results and owners carry forward. Opencontroller by Lyzr provides a governed registry of agents, versions, tools and MCP servers, policy as code, and cost attribution per agent and per version through its LLM Gateway. Opencontroller is licensed per deployment, not per seat; pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
