Rolling out AI controls to developers without stopping their work

Why are developers a special case?

Developers use coding agents that act on their machines, add MCP servers themselves, and hold access to source code, cloud accounts and sometimes production. They also have the skills and the motive to work around a control that slows them down. A rollout that treats them as the main users, rather than as a risk to be contained, tends to go better.

How should the pilot be set up?

1. Pick one team that already uses coding agents and is willing to give feedback.
2. Agree in advance which actions are in scope: package installs, shell commands, cloud CLIs, MCP servers, file access outside the project.
3. Run every rule in watch-only or simulation mode for the first one to two weeks.
4. Review the results with the team: what would have been asked or blocked, and whether that was right.
5. Switch the agreed rules to ask, and keep blocks for the short list of actions that must never run.

What should developers see?

When a control pauses an agent, the developer should see the exact command or tool call, the reason, and a way to approve, decline or ask for an exception. Products describe this step in different words: Bay returns Allow, Ask or Deny, Onyx Security lists ask and steer, and Zenity describes modifying a tool call before it executes. Ask each vendor to show you the developer's view, not only the administrator's.

How do you deliver the control?

On managed developer machines, the least disruptive path is usually the one that uses tooling already there. Bay and Noma Security state that they work through existing EDR or MDM without a new endpoint agent; Lasso Security sets Claude Code hooks through the enterprise management platform. Whatever the path, ask what changes on the developer's machine, what it does to performance, and what happens if the control is unavailable.

What should you measure?

• Interruptions per developer per day, and how many were approved.
• Time for a newly installed agent or MCP server to appear in the inventory.
• Time to answer an exception request.
• Actions blocked that the team agreed should be blocked, and any that should not have been.

When is the rollout done?

When the pilot team would not want to go back, the rules have settled, and the exception path works. Then widen to the next team, starting again in watch mode for their tools.

Keep the pilot team involved after it ends. They are the people other developers will ask whether the control is worth it, and their account of what changed will do more for the next rollout than any announcement from the security team. A short written summary of the pilot, with the measurements above and the rules that changed along the way, also gives you something to show auditors and leadership.

Sources