Four mill wheels stand around one central well, each lifting water from it and pouring water back, while a silhouetted miller at a wooden sluice decides which channel runs, the whole scene drawn with wide empty margins on every side.

Microsoft Command Line

The new primitives of agentic software delivery

Tim Bozarth · 27 October 2026

Company story · agents · artifacts · Changing roles

Takeaway

Microsoft's Tim Bozarth argues that teams should rebuild software delivery with AI agents one stage at a time, defining each stage's human job and output, rather than changing everything together.

Summary

Tim Bozarth is a corporate vice president of engineering at Microsoft. Bozarth argues that treating AI-built software as intent going in and software coming out hides the stages in between, which change at different speeds. The article sets out four stages, plan, build, verify and observe, each with a human job, an output and a link to shared knowledge.

Key points

Agents take on more implementation and testing. People choose worthwhile problems, guide consequential design decisions and judge whether the result is what the team set out to build.
Each stage hands on an output: goals and acceptance criteria, a candidate pull request with its context, auditable evidence of behaviour, and real outcomes that inform the next goals.
A shared knowledge layer links goals, decisions, evidence and production signals, keeps each item's source and version, and separates observations from interpretations.
Bozarth proposes a test for agent memory: whether an agent does a job better with the synthesised knowledge than without it.

Implication

An engineering team can list each delivery stage's inputs, human job, output and success criteria before automating further. This is a vendor's account: Bozarth describes the approach alongside Microsoft's own Copilot products and reports no measured results.

Plays

Engineering

For each stage of software delivery, write down the inputs, the jobs for people and AI agents, the output and the criteria for success, and improve the stages separately.

Engineering

Run an AI agent on the same job twice, once with its stored memory and once without, and compare speed and quality to see whether the memory helps.

Derived by Working Surface from the article; more in the Playbook. Source line: A software delivery system can execute a loop autonomously without establishing that its goal was worth pursuing in the first place.

Source issue

31 October 2026: as AI agents produce more of the work, teams move the checking to the people and records closest to the real output.