Why Ai Agents Are ‘just Another Backend’

In 2014, asking companies why they’d want to merchandise package much than erstwhile a 4th was a morganatic question. Gartner was moreover pushing “bimodal development,” advising banks to merchandise their codification only doubly per twelvemonth for stability.

“Now, if personification says, ‘My extremity is to merchandise erstwhile a quarter,’ you’re like, what are you moreover doing pinch each your time?” said Edith Harbaugh, co-founder and CEO of LaunchDarkly, successful a speech astatine AWS re:Invent.

In this section of The New Stack Makers, Harbaugh sat down pinch TNS Founder and Publisher Alex Williams to talk astir treating AI agents arsenic conscionable different backend tool, and really Harbaugh is watching complete 5,000 LaunchDarkly customers worldwide merchandise dozens of times a time while looking up to a early wherever package fixes itself.

No More ‘Golden Discs’

Today, we return CI/CD and rolling releases for granted, but we had to get present by unlearning decades of engineering practices built astir beingness constraints.

As Harbaugh said, “When you were shipping a beingness disk and users installed your package connected their ain hardware, you had to get it right. You can’t spell backmost and hole that disk.” That unit to vessel a cleanable “golden disc” shaped everything from merchandise cycles to QA processes.

Even aft package moved to nan cloud, though, bequest reasoning lingered. Users still expected to download and instal updates. It took clip for engineering practices to drawback up to nan reality that successful a Software arsenic a Service (SaaS) world, “the customer sees what nan customer sees” — nary type numbers, nary waiting for adoption.

Now AI is accelerating things moreover further, pinch nondeterministic results that tin displacement astatine immoderate moment. But Harbaugh sees continuity beneath nan disruption.

“If you’re pushing features aggregate times per week aliases moreover day, there’s a batch much travel authorities and a faster feedback loop pinch customers,” she said. “Our customers usage LaunchDarkly’s built-in A/B testing motor to trial different backends, different throughputs. In immoderate ways, an AI is conscionable a different backend.”

The Coming Age of Self-Healing Software

But Harbaugh is astir animated erstwhile talking astir what’s next: AI agents that don’t conscionable observe problems, but hole them autonomously.

“The adjacent measurement beyond detecting reliability and rolling backmost is to observe what went incorrect and person generated codification based connected nan correction fresh to cheque in,” she said.

Picture an supplier that notices failures connected a circumstantial iOS version, writes a fix, commits it and rolls it out, each pinch minimal quality intervention.

She sees akin imaginable successful automating responses to customer feedback. For example, if support tickets commencement flooding successful requesting Spanish aliases Chinese connection support, an AI supplier could grip nan localization automatically.

“Localization has historically been very manual, but that is 1 area wherever AI tin beryllium highly inexpensive and accurate,” she said.

For Harbaugh, AI fits neatly into nan ngo LaunchDarkly declared much than a decade ago: launch, measure, control. “For us, an supplier is fundamentally conscionable different group of software. You’re still measuring consequence time, accuracy and expected outcomes,” she said. “Agentic AI tin trigger overmuch faster, but you interest astir each nan aforesaid things.”

The afloat speech covers really LaunchDarkly’s AI configs thief customers trial different ample connection models (LLMs) against business metrics, nan situation of versioning information models during unrecorded releases and why much AI codification agents won’t mean less engineers, conscionable for illustration moving from nan descent norm to spreadsheets didn’t destruct Wall Street jobs.

Group Created pinch Sketch.