---
title: How do you govern AI agents built by citizen developers?
url: https://opencontroller.lyzr.com/blog/govern-ai-agents-built-by-citizen-developers
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do you govern AI agents built by citizen developers?

**Govern citizen-built AI agents with defaults, not reviews. Give makers a safe sandbox environment, restrict which connectors and data each environment can use, cap how widely an agent can be shared, and require a named owner. Agents that need wider reach, sensitive data or write actions move through a defined promotion path to IT.**

Low-code builders put agent creation in the hands of analysts, sales ops and HR teams. That is the point of them. It also means the number of agents grows faster than any review board can read. The goal is not to stop makers. It is to make the safe path the easy path.

## Why can't you govern citizen agents the same way as pro-code agents?

Pro-code agents pass through a repository, a pipeline and a release process. Citizen agents usually do not. They are built in a browser, saved in a tenant and shared with a link.

Three things change:

- **Volume.** A large tenant can hold hundreds or thousands of maker-built agents. Case-by-case review does not scale.
- **Visibility.** There is no pull request to inspect. The agent's configuration lives inside the platform.
- **Ownership.** Makers change roles and leave. Orphaned agents keep running on their credentials.

So the controls must sit in the platform configuration and in an inventory, not in a manual approval queue.

## What should the guardrails-by-default checklist include?

Use this as a baseline. Each item is a setting or a process, not a policy statement.

1. **Default environment is a sandbox.** New agents land in an environment with no production connectors and no external sharing.
2. **Data policies per environment.** Group connectors into allowed and blocked sets. Block consumer storage, personal email and unknown HTTP connectors in maker environments.
3. **Maker credentials off for shared agents.** An agent shared with others should authenticate as the end user or a managed identity, not as the maker.
4. **Sharing limits.** Cap the number of people or groups an agent can be shared with until it is promoted.
5. **Named owner and backup owner.** Required at creation. Re-confirmed when either person leaves.
6. **Minimum registration fields.** Purpose, data touched, actions it can take, audience. Keep it to five or six fields so makers fill it in.
7. **Write actions need promotion.** Read-only agents can stay in the maker tier. Agents that send email, update records or call external APIs cannot.
8. **Automatic inventory.** Discover agents from the platform's admin APIs on a schedule. Do not rely on self-reporting alone.
9. **Inactivity expiry.** Flag agents with no use in 60 or 90 days. Archive them after notice.
10. **Training before building.** A short course on data handling and prompt injection before maker licences are assigned.

## How should the promotion path from maker to IT work?

Define tiers so makers know what triggers a review.

| Tier | Who can use it | Data and actions | Review |
|---|---|---|---|
| Personal | The maker only | Maker's own data, read-only | None beyond defaults |
| Team | Up to a set group size | Approved internal sources, read-only | Owner attests registration fields |
| Department | Wider internal audience | Internal data, limited write actions | Risk tiering and an evaluation run |
| Enterprise or external | All staff or customers | Sensitive data or external actions | Full production readiness gate, IT co-ownership |

The trigger to move up is usage or capability, not the maker's request. If an agent's sharing count or connector list crosses a tier boundary, the platform setting should block it until the review is done.

See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents) and the [production readiness gate](/blog/production-readiness-gate-for-ai-agents) for the higher tiers.

## Which platform settings do the heavy lifting?

Most low-code agent platforms expose the same levers under different names. On Microsoft Power Platform, for example, these are environments, [data policies](https://learn.microsoft.com/en-us/power-platform/admin/wp-data-loss-prevention) that classify connectors as business, non-business or blocked, and [Managed Environments](https://learn.microsoft.com/en-us/power-platform/admin/managed-environment-overview) that add sharing limits and usage insights. We cover the specifics in [Copilot Studio data policies for agents](/blog/copilot-studio-data-policies-for-agents) and [limiting who can share Copilot Studio agents](/blog/limit-who-can-share-copilot-studio-agents).

The pattern carries across platforms: separate environments, connector allow-lists, sharing caps and admin-API discovery.

## What goes wrong most often?

- **Everything in the default environment.** Makers build where they land. If the default allows production connectors, so does every agent.
- **Inventory by spreadsheet.** Self-reported lists are out of date within weeks. This is how [AI agent sprawl](/blog/what-is-ai-agent-sprawl) starts.
- **Reviews with no exit.** If promotion takes months, makers share links informally instead.
- **No offboarding hook.** The HR leaver process does not reassign or retire the person's agents.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to third-party agent platforms and Claude Enterprise and OpenAI Enterprise workspaces. Discovered agents land in a governed registry with owners, versions, tools and MCP servers. Posture flags risky permissions and configuration, and per-agent Block and Quarantine controls act through the connected platform. Depth of control depends on the connector and traffic path. See the [capability reference](/capability-reference) or the pillar on [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

To see this on your own estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should citizen developers be allowed to build AI agents at all?**
Yes, with defaults. Blocking low-code tools pushes makers to unsanctioned tools with less visibility. A sandbox environment, connector limits and sharing caps keep risk low while allowing experimentation.

**Who owns a citizen-built agent after the maker leaves?**
The named backup owner, then the maker's manager. Tie agent ownership to the leaver process so every agent is reassigned or retired.

**How many registration fields should makers complete?**
Five or six: purpose, owner, data sources, actions, audience and business unit. More fields reduce completion rates without adding much risk signal.

## Key facts

Citizen-built AI agents are best governed through platform defaults: sandbox environments, connector data policies, sharing limits, named owners and a tiered promotion path. Opencontroller discovers agents across clouds and connected platforms, keeps them in a governed registry, flags risky configuration with Posture and can Block or Quarantine agents through the connected platform. Opencontroller is licensed per deployment per year, starting from $250K.

## Related

- [What is AI agent sprawl and how do you contain it?](https://opencontroller.lyzr.com/blog/what-is-ai-agent-sprawl.md)
- [How do Power Platform data policies control Copilot Studio agents?](https://opencontroller.lyzr.com/blog/copilot-studio-data-policies-for-agents.md)
- [How do you limit who can share Copilot Studio agents?](https://opencontroller.lyzr.com/blog/limit-who-can-share-copilot-studio-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
