How do you plan a percentage rollout?

What steps should a rollout go through?

There's no single right schedule. Flagger.gg's guide to percentage rollouts describes a pattern many teams use:

1. Admins first, so your own team sees it before any customer.
2. A small share, such as 5% or 10%, held long enough for errors and support tickets to show up. For a busy endpoint that may be an hour; for a weekly workflow, days.
3. 25% and 50%, where load problems appear that a small share hides.
4. Everyone, then remove the flag and the old code path.

What should make you stop?

Decide before you start: an error rate, a latency number, or a metric that mustn't drop. Write it down with the flag, so whoever is watching, a person or an agent, knows the line. In Flagger.gg, watching those numbers is up to your own monitoring for now; tying a rollout to metrics, with automatic rollback, is listed as in development.

Why does sticky bucketing matter?

If users were picked at random on each request, someone could see a feature appear and disappear. Sticky bucketing gives each user a fixed bucket per flag, so raising the percentage only adds people and lowering it removes the most recently added first. Flagger.gg computes the bucket from a hash of the flag key and the user id, so the first 10% isn't the same 10% on every rollout. Anonymous users stay off until the flag is fully on.

How do you roll back?

• Narrow it: lower the percentage. Making a flag narrower never needs approval.
• Kill it: turn it off for everyone at once.
• Revert it: put the flag back to its state before the latest change.

flagger set new_search percent --percent 5 --note "errors at 25%"
flagger kill new_search --note "checkout errors"
flagger revert new_search

Each reaches your servers within a second, because the check runs in your app from a streamed ruleset. See how to stop an AI agent that's going wrong.

Can an AI agent run the rollout?

Yes, within limits a person sets. Over MCP or the CLI, an agent with the rollout action raises the percentage step by step, with a note on each change. In Flagger.gg's auto-live mode, the default, there's no cap. With approvals on, an agent in production can go up to the environment's cap, 10% by default, and anything wider becomes a request a person approves; the CLI exits with code 2 when a change is waiting. See how to add human approval to an AI agent's actions and what a feature flag is.