Empirical demonstration: does architecture that is good for humans also make coding agents more reliable?

This is a research artifact, not a framework. It packages two functionally equivalent TypeScript apps, a frozen task suite, instruction-pack conditions, a thin harness, and a case study template designed to falsify a comfortable prior—not market it.

The characteristics that make a codebase easier for humans to maintain — consistency, small abstractions, explicit contracts, strong tests, predictable organization, limited degrees of freedom — also make coding agents materially more reliable.

The author’s belief is that “agent-friendly architecture” is mostly classical software engineering. The experiment is built to allow that belief to fail.

Primary agent: Claude Code. Optional Codex adapter is non-primary.

pnpm install
pnpm test:apps          # both apps + product acceptance
pnpm harness -- --app canonical --instructions router --task T01 --dry-run
./demo/reproduce.sh

Run a live agent task (requires claude CLI):

pnpm harness -- --app canonical --instructions router --task T01 --live

Ideas at the application layer mirror process patterns from:

• mwarger/ralph-crispies — small units, gates, fresh context
• mwarger/forge — hypothesis → act → evaluate; blind judging
• mwarger/super-ralph — vertical slices, fat specs, dumb runners

Here the question is not how to orchestrate agents, but whether the codebase under edit changes agent reliability.

Read METHODOLOGY.md before changing tasks or rubrics. Live experimental runs should not begin until methodology is tagged/committed.

• Methodology, product contract, rubric, tasks
• Canonical + heterogeneous OpsDesk apps
• Product acceptance parity suite
• Instruction packs (giant / router)
• Thin harness + demo scaffold
• Pilot live runs (T01, T03, T04 × 4 cells)
• Full suite + filled CASE_STUDY.md

Commits are authored solely as Mat Warger <686823+mwarger@users.noreply.github.com>. Verify with ./scripts/verify-authors.sh.

MIT