What Happened
Birgitta Böckeler, a Distinguished Engineer at Thoughtworks, published a new article on Martin Fowler's site introducing the concept of "harness engineering" - the practice of building tooling and constraints around AI coding agents to keep them producing quality code at scale.
The article breaks the harness into three components. First, context engineering: giving agents access to enhanced knowledge bases within codebases, observability data, and browser navigation so they understand what they're working with. Second, architectural constraints: deterministic custom linters, structural tests (referencing frameworks like ArchUnit), and enforced module boundaries that catch problems in AI-generated code before it ships. Third, garbage collection: periodic agents that scan for documentation inconsistencies and architectural violations that accumulate over time.
The piece references OpenAI's Codex team as a case study, noting they produced over 1 million lines of code without manual typing over five months using their own harness setup. Böckeler flags a notable gap in that example: there's no verification of actual functionality and behavioral correctness in the reported results.
The article recommends teams start by auditing their current harness - examining pre-commit hooks, considering custom linters, identifying architectural constraints worth enforcing, and experimenting with structural testing frameworks.
Why It Matters
If you're using AI coding tools like Cursor, Claude Code, or Aider, this directly affects how much you can trust what they produce. Right now, most developers rely on vibes - reading through generated code and hoping it looks right. That doesn't scale.
The harness approach formalizes what effective AI-assisted teams are already doing informally: setting up guardrails that catch bad output automatically. Pre-commit hooks that reject code violating architectural patterns. Linters that enforce naming conventions and module boundaries. Structural tests that verify the shape of the codebase hasn't drifted.
This matters most for teams scaling AI code generation beyond one developer's side project. When multiple agents are generating code across a large codebase, the compounding effect of small inconsistencies gets serious fast.
Our Take
The framing here is useful. "Harness engineering" gives a name to something the AI coding community has been circling around without clear vocabulary. The three-part breakdown - context, constraints, garbage collection - maps well to what we've seen work in practice.
The OpenAI Codex case study is the weakest part. A million lines of code is a vanity metric without knowing how much of it actually works correctly. Böckeler rightly calls this out, and it's the central tension of AI-assisted coding right now: output volume is easy, output quality is hard.
The practical takeaway is clear. If you're using AI coding assistants seriously, invest in your harness before you invest in more agent capabilities. A well-configured set of linters, pre-commit hooks, and structural tests will do more for your code quality than switching to a fancier model. The agents are already good enough to produce volume - the bottleneck is verification.