Software Factories
Details
Overview
A taster for our ML engineering programme. You'll learn what it takes to leave a coding agent running for hours, and then design a software factory, where people write what to build and how to check it and agents write the code in a loop until the checks pass.
Background
To leave a coding agent running for hours, the harness has to keep the cost down, keep the context small and check the work. Each call sends the whole context to the model again, so over a long run the same tokens are processed many times. Prompt caching brings the cost of those tokens down, and it's built on the KV cache the model keeps while it generates.
As the context grows, the model does worse on the same task, and this starts well before the context window is full. Harnesses keep the context small with compaction, sub-agents and memory files. A Ralph loop runs the agent on the same prompt again and again, each time in a fresh context, and what earlier runs did stays on disk in the plan and the git history.
A coding agent can also stop before the work is done, and say a feature works when it was never run. Verification is a check the harness runs to find out whether the work is done, and the loop ends when the check passes. In a software factory these checks do the job that code review did. We'll look at the factory StrongDM's AI team runs, in which code is neither written nor reviewed by people.
Who is this for
Experienced software engineers who have used a coding agent such as Claude Code, Codex or Cursor. You don't need an ML background.
What we will cover
- Opening talk: KV caching, prompt caching, context rot, verification and the Ralph loop
- Software factories, with StrongDM's as an example
- Small groups at the whiteboard: choose a project and design a software factory to build it
- Start building your factory with your own coding agent
What to bring
- Your laptop
- A coding agent you already use, such as Claude Code, Codex or OpenCode
