Article —
Microsoft maps agentic delivery to jobs, artifacts and memory
Tim Bozarth of Microsoft argues that agentic delivery needs more than code generation. It needs defined jobs, artifacts and a knowledge layer that lets later runs learn from earlier ones.
The new primitives of agentic software delivery is an essay by Tim Bozarth on Microsoft’s Command Line blog. It argues that AI-first development goes beyond code generation and has to connect people, AI and feedback across the whole software lifecycle.
The argument
Describing a goal and getting software back treats delivery as a black box. The essay says each stage of the lifecycle changes at its own pace, and bottlenecks move as tools improve. Changing everything at once makes it hard to see what works.
Its advice is to start with three things:
- The job each stage must do, and which part of it stays human.
- The artifact each stage hands to the next.
- The knowledge that connects the stages.
The stages are Plan, Build, Verify and Observe. Humans move toward judgment about value, intent and customer experience. Agents take more of the implementation. In the author’s words: “Autonomy does not remove that structure.”
The knowledge layer holds goals, design decisions, evidence, release history, production signals and agent behaviour logs. It tracks where each item came from and keeps observations apart from interpretations. On top sits memory: synthesised lessons that help an agent avoid repeated mistakes. The test is simple. Does the agent do better with it than without it?
The essay points to GitHub Copilot’s coding agent, Copilot Code Review, MCP and agent skills as examples of the pieces.
What we take from it
The record of earlier runs is the point. Without it, every run starts cold. With it, each run can improve the next. That is the most useful idea in the text, and it is stated clearly.
Where we would want more
The essay does not say how to store this knowledge or which technologies to use. A reader can imagine it: a database, an index, a log of runs. But a concrete example would make the idea easier to act on. Treat the article as a map of what to build, not a guide to building it.
Who should read it
Teams that already run coding agents and wonder why results do not improve over time. It gives a shared vocabulary for stages, hand-offs and memory. It is also a fair pair to other agent tooling we have covered: same questions, different answers.
Contains AI-generated content