← Back to posts

Series

Build an AI software factory (an experiment in .NET)

An AI software factory, an experiment in .NET: hand off the work, keep the judgement. A pipeline run of agents and checks, a Start workflow dialog for a brief, and JetBrains Rider with review notes as cards next to the code, including a question answered by Claude
A series about handing software work to AI coding agents without handing off your judgement. I'm building a factory around them in .NET: where the work comes from, where agents run, how their output gets checked, what it costs, what they remember, and how people stay in charge of what ships. Each part takes one piece and describes what I learn while building it.
1 part 58 min total read
  1. Part 1

    What it takes to build an AI software factory (an experiment in .NET)

    An AI coding agent, whether it lives in your terminal, Visual Studio, Rider or VS Code, has a safety system you rarely think about: you. You approve the commands, notice when it drifts, stop it when it burns money, and remember what it learned yesterday. A software factory is what you get when you take yourself out of the room, and every one of those jobs has to become a part of the system instead. This post is my best guess at that parts list, based on what I've learned from experimenting so far.

    Sep 15, 2026

Planned posts

  1. Sandboxes: containers, worktrees, egress allowlists and a credential proxy, so an agent can run without permission prompts.
  2. Pipelines: phases, typed hand-offs, tool boundaries per step, and why code owns the sequence.
  3. Gates: accepting work on evidence the agent didn’t write, and sending failures back to the right session.
  4. Budgets and metrics: spending limits enforced by the spender, pricing transcripts correctly, and cost per surviving line of code.
  5. Memory: pushed facts, pulled knowledge, and why nothing becomes memory without a person confirming it.
  6. Skills: scoping, sourcing and versioning skills across a pipeline.
  7. Harnesses: running Claude Code, Codex, Copilot CLI, Junie CLI and others behind one contract, and comparing them fairly.
  8. Review: a review tool built for agent output, in the IDE, the terminal or the browser: files grouped by kind, comments with a scope and a kind, and a review that goes back to the agent as its next brief.
  9. Fleets of agents: a planner that splits the work, checks on the division before a fleet starts, and a board for agents in separate sandboxes to talk on.
  10. The feedback loop: judging finished runs and improving the factory without letting it rewrite its own rules.

These follow one by one after the introduction above, and the order may shift as they get written.