In 2014, asking companies why they’d want to release software more than once a quarter was a legitimate question. Gartner was even pushing “bimodal development,” advising banks to release their code...
In 2014, asking companies why they’d want to release software more than once a quarter was a legitimate question. Gartner was even pushing “bimodal development,” advising banks to release their code only twice per year for stability.
“Now, if someone says, ‘My goal is to release once a quarter,’ you’re like, what are you even doing with all your time?” said Edith Harbaugh, co-founder and CEO of LaunchDarkly, in a conversation at AWS re:Invent.
In this episode of The New Stack Makers, Harbaugh sat down with TNS Founder and Publisher Alex Williams to talk about treating AI agents as just another backend tool, and how Harbaugh is watching over 5,000 LaunchDarkly customers worldwide release dozens of times a day while looking ahead to a future where software fixes itself.
No More ‘Golden Discs’
Today, we take CI/CD and rolling releases for granted, but we had to get here by unlearning decades of engineering practices built around physical constraints.
As Harbaugh said, “When you were shipping a physical disk and users installed your software on their own hardware, you had to get it right. You can’t go back and fix that disk.” That pressure to ship a perfect “golden disc” shaped everything from release cycles to QA processes.
Even after software moved to the cloud, though, legacy thinking lingered. Users still expected to download and install updates. It took time for engineering practices to catch up to the reality that in a Software as a Service (SaaS) world, “the customer sees what the customer sees” — no version numbers, no waiting for adoption.
Now AI is accelerating things even further, with nondeterministic results that can shift at any moment. But Harbaugh sees continuity beneath the disruption.
“If you’re pushing features multiple times per week or even day, there’s a lot more flow state and a faster feedback loop with customers,” she said. “Our customers use LaunchDarkly’s built-in A/B testing engine to test different backends, different throughputs. In some ways, an AI is just a different backend.”
The Coming Age of Self-Healing Software
But Harbaugh is most animated when talking about what’s next: AI agents that don’t just detect problems, but fix them autonomously.
“The next step beyond detecting reliability and rolling back is to detect what went wrong and have generated code based on the error ready to check in,” she said.
Picture an agent that notices failures on a specific iOS version, writes a fix, commits it and rolls it out, all with minimal human intervention.
She sees similar potential in automating responses to customer feedback. For example, if support tickets start flooding in requesting Spanish or Chinese language support, an AI agent could handle the localization automatically.
“Localization has historically been very manual, but that is one area where AI can be extremely cheap and accurate,” she said.
For Harbaugh, AI fits neatly into the mission LaunchDarkly declared more than a decade ago: launch, measure, control. “For us, an agent is basically just another set of software. You’re still measuring response time, accuracy and expected outcomes,” she said. “Agentic AI can trigger much faster, but you worry about all the same things.”
The full conversation covers how LaunchDarkly’s AI configs help customers test different large language models (LLMs) against business metrics, the challenge of versioning data models during live releases and why more AI code agents won’t mean fewer engineers, just like moving from the slide rule to spreadsheets didn’t eliminate Wall Street jobs.
The post Why AI Agents Are ‘Just Another Backend’ appeared first on The New Stack.
