Guides and Sensors

Harness controls are feedforward guides (rules, docs) or feedback sensors (tests, analysis, reviewers), each computational or inferential; sensors are placed by cost, cheap ones locally and fuller ones later.

Evidence: moderate

A harness accumulates many kinds of control: instruction files, linters, tests, review agents, scheduled checks. Without a way to classify them it is hard to see what is missing, what duplicates what, or where an expensive check is running too often.

Birgitta Böckeler's framing, from her martinfowler.com article (read through a secondary summary) and confirmed in her own words on SE Radio 730, sorts every control along two axes. A guide acts before the agent does: AGENTS.md files, rules, and computational guides such as codemods and OpenRewrite recipes. A sensor acts after: tests, static analysis, and large language model (LLM) reviewers. Each guide or sensor is either computational (deterministic code) or inferential (judged by a model). Sensors are then placed by cost, with cheap ones run locally, fuller ones at pre-commit, and maintenance sensors on a weekly schedule. She also introduces "harnessability", the observation that some codebases are easier to harness than others, and advises assessing risk as probability times impact times detectability while keeping humans accountable.

Böckeler is candid about open problems: measuring how effective sensors are, conflicting constraints, guides going stale as models improve, and the sprawl of dozens of markdown files.

The frame explains several neighbouring practices. Deterministic sensors are what back-pressure means in practice, a short steering file that points rather than explains is a guide, and fix the environment, not the output amounts to adding a guide or sensor after every failure. An LLM reviewer is an inferential sensor in the role of a separate evaluator.

For an agent that diagnoses or operates live systems, the frame suggests an ordering, offered as an implication rather than a sourced claim: health checks, status commands, logs and diagnostic tools are computational sensors and should run before any LLM is involved. That is the same ordering argued in deterministic scores gate LLM diagnosis.

Sources

Software Engineering Radio, July 2026 — https://se-radio.net/2026/07/se-radio-730-birgitta-boeckeler-on-harness-engineering-for-ai-agents/
Birgitta Böckeler, Feb 2026, read via secondary summary — https://github.com/robyscar/ML_LLM_CLAUDE_SKILLS_Jamie-BitFlight_claude_skills/blob/main/research/evaluation-testing/harness-engineering-martin-fowler.md