Jyoti Bansal on how harness is rethinking AI for software delivery

Every developer tool company is racing to add AI features. Few are rethinking the entire delivery pipeline from first principles. Jyoti Bansal, founder and CEO of Harness, is one of the people doing exactly that — and he's been building developer infrastructure long before the current wave.

WorkOS CEO Michael Grinich caught up with Bansal at HumanX 2026 in San Francisco to talk about AI-native software delivery, the real bottlenecks in developer productivity, and what changes when AI agents start shipping code.

From AppDynamics to Harness

Bansal founded AppDynamics and built it to a point where Cisco acquired it for $3.7 billion in 2017, just before its planned IPO. With Harness, he turned his attention to the other side of the software lifecycle: delivery. The thesis was straightforward — observability had been modernized, but CI/CD and the broader delivery pipeline were still dominated by tools like Jenkins that required heavy manual configuration.

Harness set out to build a modern, AI-native platform for software delivery — covering continuous integration, continuous deployment, feature flags, cloud cost management, and more. These aren't separate problems. They're all part of getting code from a developer's machine into production reliably.

AI changes the delivery problem

The conversation zeroed in on what happens to software delivery when AI starts generating a significant share of production code. Bansal's take: the bottleneck shifts. When developers can write code faster — or when AI agents write it for them — the constraint moves downstream. Testing, security scanning, deployment verification, and rollback automation all become more critical, not less.

This isn't a hypothetical. Harness is already seeing customers where AI-generated code volume is increasing. The pipeline has to keep up. That means smarter test selection, faster feedback loops, and delivery systems that can reason about risk rather than just execute a static YAML config.

The AI-native platform bet

Bansal drew a clear line between "AI-assisted" and "AI-native." Bolting a chatbot onto an existing CI/CD tool doesn't change the fundamental architecture. An AI-native delivery platform, in his framing, means the system itself uses AI to make decisions — which tests to run based on code change analysis, how to canary a deployment by gradually shifting traffic and monitoring error rates, when to auto-rollback based on real-time health signals, and how to optimize cloud spend by identifying underutilized resources.

Rather than treating AI as a feature, Harness is embedding intelligence into the platform's core decision-making. The goal: a delivery pipeline that adapts based on what it learns from every deployment, not one that requires a human to hand-tune every stage.

What developers actually need

One of the sharper points in the conversation: developer productivity isn't just about writing code faster. The industry has a tendency to over-index on the code generation piece — how many lines of code can an AI produce per hour. Bansal argued that the real productivity gains come from reducing the time between writing code and knowing it works in production.

That means faster builds, smarter test suites that use code change impact analysis to skip irrelevant tests, deployment strategies like canary and blue-green that reduce risk without slowing down releases, and cost visibility so teams aren't burning money on idle infrastructure. These are the unsexy parts of the developer experience, but they're where teams lose the most time.

Where this is headed

Bansal sees a near future where AI agents are full participants in the delivery pipeline — not just generating code, but also reviewing it, testing it, deploying it, and monitoring it in production. The human role shifts toward oversight, architecture decisions, and handling the edge cases that AI can't yet reason about.

Platforms that can orchestrate both human and AI contributors across the full delivery lifecycle will have a structural advantage. Static pipelines defined entirely in YAML won't cut it when a growing share of commits are coming from AI agents that work at machine speed.

This interview was recorded at HumanX 2026 in San Francisco.