Today we're introducing Release Autopilot, new automation on the LaunchDarkly platform that gives your agents the power to safeguard changes all the way to production.
It handles the release work engineers have had to do by hand: identifying risky changes, adding a flag where needed, rolling them out progressively, rolling them back if something degrades, and cleaning up afterward. When release work is automated, teams can safely release the growing volume of code they're building with AI.
This has changed how our team works. Letting agents release features without human review used to be unthinkable, but today at LaunchDarkly they already ship over 20% of pull requests directly to production. Engineers set the guardrails, agents handle the rest. It’s a big part of how we’ve scaled AI-led development.
The factory doesn’t end at deploy
Every engineering team is working through the same question right now: What happens to the software development lifecycle when you put ten times as much code through it? The massive acceleration in development speed has pushed the bottleneck downstream to validation, deployment, and release.
We’ve been living this ourselves and learning from our customers. The natural response is to invest in code quality, with better harnesses, more context, more tests, and agentic review. We're doing all of that at LaunchDarkly.
But quality is only half of the equation. Tests cover the scenarios we anticipate, and code review catches the problems we can identify. That’s true whether a human or agent is doing the work. Production is where real customers show up, and with them come edge cases and traffic patterns no one fully anticipated.
We have to invest in the second half of the equation as well: reducing the risk of what ships in production. If you can control who sees a change, measure it on its own, and turn it off instantly when something is wrong, you can limit the blast radius of any mistake—and feel confident moving at agent speed.
Reducing risk in production is what LaunchDarkly has done for more than a decade. What’s different now is the volume. With ten times the volume of change, relying on an engineer to create a flag, choose rollout steps, select metrics, manage a multi-phase release process, and watch a dashboard makes release a bottleneck.
We know what good releases should look like. But humans often take shortcuts because of the toil and friction involved in following established best practices. The answer is to let agents automate the repetitive steps to deliver great releases every time.
What we're launching
Release Autopilot is automated release control for AI-built code. It encodes the best practices we've been recommending for years and applies them without asking anyone to remember. When a change lands, Release Autopilot can:
- Agentically assess the pull request and classify its risk.
- Update the pull request to add a feature flag when the change needs one.
- Roll the change out progressively, watching the guardrail metrics you define.
- Hold or roll back automatically if a metric moves the wrong way.
- Submit a pull request that removes the old code and archive the flag once the release succeeds.
Control, measurement, and response, on every change, with no one babysitting it.
Release Autopilot isn’t a code review tool, and it doesn't replace your CI/CD pipeline. CI/CD automates how code gets deployed. Release Autopilot automates how a change reaches your users once it's deployed. It works alongside the agents and pipelines you already use and runs on the same LaunchDarkly platform that evaluates 70 trillion flags a day.
You decide how much autonomy to give the system. Teams can start with recommendations and human approval, then expand automation as trust builds. At LaunchDarkly, we fully automate low-risk changes, while engineers review higher-risk ones.
Valuable software per hour of human attention
Even at LaunchDarkly, where we have championed feature management for more than a decade, we’ve had room to improve. For a lot of changes, the cost-benefit of adding a flag wasn’t worth it, except when it was and we didn’t know it. Now with Release Autopilot, automation removed cost from that decision.
Over the past year, with roughly the same number of engineers, we’ve more than doubled monthly merged PRs and nearly quadrupled feature flag creation, with automation now handling 42% of flag removals. Despite the massive increase in the volume of change, our Major+Critical incident time has decreased by about 70%
When we started building our AI software factory, we set a goal: Maximize how much valuable software reaches customers for each hour of human attention it takes to get there. Lines of code, tokens consumed, and pull requests opened don’t tell you that. The point of all this automation is to let engineers deliver more to customers and spend less time managing the machinery.
Thank you to the teams who have tested Release Autopilot with us and to every customer who has trusted LaunchDarkly with their releases over the past decade.
Check out Release Autopilot today. Learn more, start a trial, or talk to us about how this fits in your factory here. Try it free today. GA Nov 16; certain features in beta.
–Jonathan
