ADLC Is the New DevOps for AI Agents — Here's What That Means | Blog

DevOps was built for deterministic systems — code ships, tests pass, infrastructure scales, alerts fire. It’s a tight feedback loop designed around predictable execution. But when the primary unit of execution is an AI agent that reasons, defers, calls external APIs, and makes autonomous decisions over hours or days, that loop breaks. A hallucination isn’t a failed test; it’s a live action in production. A state loss isn’t a bug in a log; it’s a process that quietly stopped doing what it was supposed to do. That’s the problem ADLC was built to solve.

What Is ADLC?

The Agent Development Life Cycle (ADLC) is a structured methodology for building, deploying, and continuously governing AI agents. It’s not DevOps with an AI-themed makeover — it’s a different operating model that accepts non-determinism, evolving behavior, and persistent state as core architectural realities rather than edge cases to be managed around.

Where DevOps assumes code does what you wrote, ADLC accounts for the fact that agent behavior depends on prompts, context windows, model versions, and external tool availability — any of which can shift the outcome without a single line of code changing. IBM, Arthur AI, EPAM, and Cycode have all begun publishing ADLC frameworks in the past several months, each converging on a similar set of phases.

Why DevOps Falls Short for Agents

The gap becomes clearest in two areas: state management and governance.

Traditional DevOps pipelines treat services as stateless where possible and rely on infrastructure-level restarts when something breaks. Persistent agents can’t be restarted without consequences — they carry memory, context, and partially-completed workflows. Microsoft researchers noted earlier this month that agent performance degrades while on a continuous task, and that doubling task duration can increase failure rates. That’s not a deployment problem; it’s a lifecycle problem.

On governance, the stakes are higher too. An agentic AI system with write access to external systems — databases, APIs, ticketing platforms — turns a hallucination into a potential security or operational incident, not just a user-facing typo. The higher the blast radius of the agent’s permitted actions, the higher the stakes of a bad inference.” ADLC bakes governance gates directly into each phase rather than treating security as a post-deployment concern.

Glean’s 7-Stage Framework (May 2026)

On May 12, Glean formally introduced its Enterprise ADLC: a seven-stage lifecycle spanning Opportunity, Design, Performance, Context, Develop, Launch, and Monitor & Improve. The framework ships alongside new platform capabilities including an Agent Sandbox, Content and Scheduled Triggers, and Agent Access Policies designed to close the gap between building an agent and safely running it at scale.

Glean’s CPO Emrecan Dogan put it plainly: “Agents are software. They need a disciplined way to be defined, built, launched, governed, and improved over time.” That’s the ADLC pitch in one sentence.

Use Case: Enterprise IT Ops Agent

Consider a persistent IT operations agent tasked with monitoring infrastructure alerts, triaging incidents, and filing tickets — running continuously across shifts. Under traditional DevOps, you’d instrument it like a micro-service: CI/CD pipeline, uptime monitoring, error alerting. But what happens when the agent misclassifies a P1 incident as noise at 2am because the context window drifted? There’s no failing test. No red build. Just a missed page.

Under ADLC, this agent has a defined evaluation harness tied to the Monitor phase — escalation accuracy benchmarked continuously against real outcomes, model drift detection, and a human-in-the-loop gate for any action above a defined blast radius. The “improve” phase feeds real incident data back into the agent’s evaluation suite before the next deploy window. The agent doesn’t just run — it has a lifecycle.

What Teams Should Do Now

ADLC maturity is still early, but the window for getting ahead of it is closing fast. Teams deploying anything beyond single-turn AI workflows should audit their current pipelines against ADLC phases — particularly around evaluation continuity, state persistence, and governance checkpoints at deployment. The organizations that treat agents like software with a lifecycle — not like features with a ship date — will be the ones scaling safely in 2027.

AI Disclosure

This document is drafted by an AI skill and is provided for informational and governance support purposes only. It does not constitute legal advice or a formal compliance determination. Do not publish or rely on this notice as a substitute for review by qualified legal counsel or a licensed compliance professional with jurisdiction-specific expertise.