What is a feature flag?

A feature flag is a setting that turns a piece of functionality on or off, or limits it to certain users, without releasing new code. For AI agents, flags can switch individual capabilities, sources or channels on and off. They are useful for rollout, but a flag is an off switch, not evidence that the agent may act.

Why do feature flags matter for agents?

Flags separate deploying a change from switching it on. You can ship a new capability dark, turn it on for one team, watch it, and turn it off again in seconds if something looks wrong. For agents, where behaviour on live work is hard to predict from tests alone, that control is valuable.

They also carry a risk that is easy to miss. If a flag is the only thing between an agent and an action, then whoever can flip it can grant the agent authority. A flag set to "on" says the capability is available. It says nothing about whether the agent has proven it can do the job, at what depth, or within which limits.

How do feature flags work for agents in practice?

Common uses:

• Staged rollout: a new capability is enabled for one branch, one product or a small share of cases first.
• Channel control: an agent may answer by chat but not yet by email.
• Source switching: a new source is added behind a flag so it can be turned off if it misleads the agent.
• Fast disable: a capability is turned off during an incident without a release.

Each flag should have an owner, a reason, a review date and a record of every change. Old flags left in place become hidden configuration nobody understands.

What does an example look like?

A UK business bank's operations team adds a capability that lets its servicing agent draft responses to requests for changes to a company's authorised signatories. The capability ships behind a flag, enabled for one relationship team. The agent is certified at depth 2: it drafts, a servicing officer checks the mandate and sends. After a month without issues, the flag is widened to all teams. The flag controlled who got the capability; the certificate controlled what the agent could do with it. Work of this kind is described in AI agents in UK business banking operations.

How is a feature flag different from depth?

A flag is binary and set by whoever controls configuration. Depth is graded, needed by each action, and earned by the agent through certification on the customer's own cases. Turning a flag on cannot lift an agent from drafting to executing; only evidence can do that. Turning a flag off, though, can stop a capability completely, which makes flags a reasonable emergency control alongside a proper kill switch.

If a flag enables a capability that needs more depth than the agent is certified to, the minimum chain still holds or refuses the action and names the certificate as the deciding bound. A flag that keeps producing refusals means the rollout and certification plans are out of step.

Where do flags help in regulated lending?

In consumer lending, flags are often used to restrict new agent capabilities to lower-risk products first, such as small fixed-sum loans before credit cards. That keeps early exposure to cases where the rules are simplest and the amounts smallest. The role of operations teams in that setting is set out in AI agents in UK consumer lending operations.

Which terms sit close to it?

Canary release: a flag is often how the canary's share is controlled. Kill switch: a dedicated, owned stop, which a flag can implement but should not replace. Change management: flag changes are changes and should be recorded. See our guide to deploying AI agents for the wider release process.