Agent DLC is a test method for AI agents before release

The Agent Development Lifecycle, called Agent DLC, is a method to select if an AI agent is prepared for production. The authors show it with a test of an HR policy assistant that gives answers from documents. They say the test is not a customer result and not a benchmark. Business owners give approval for one written standard, and each change is judged against it.

How it works

Agent DLC has six tasks: Define, Build, Evaluate, Release, Observe and Feed Back. All of them use one standard of criteria, limits and approved examples. Agents need more checks, because the same input can give different outputs. Small changes to prompts or settings change the behavior.

The numbers

The test gave three questions five measures each. Chunk-level precision asks if a passage is on topic. It was 1.00 for all three questions. Fact-level precision asks if each fact is necessary. It was 0.09 for the third question, because the documents had no sick-leave policy. One answer had a score of 0.90 for groundedness, but it gave an incorrect claim about the user.

What is not known

Scores change between runs. A borderline pass can become a failure. The authors say the example gives no general rate of improvement. They say it does not validate a release to production.

Insight: A passing score can hide an incorrect claim

ElevenLabs and Vercel Labs released agent tools with no test results. This test shows that a score of 0.90 can pass an incorrect claim. It looks like a race to release agent tools while proof falls behind, so the owner of the written standard decides what passes. Leaders must ask for that standard before they buy an agent.