Home

harness-kit Wiki

The knowledge base for harness-kit: a set of Claude Code agents that carry a raw idea to a merged PR through one gated pipeline, plus a system-architect agent that turns a problem into a rigorous System Design Doc.

The repo holds the condensed, agent-facing form (guides, sensors, evals as terse markdown). This wiki is the human-facing depth: the theory behind the harness, how each control works, and full study material with references for each system design.

The harness

How and why harness-kit is built the way it is.

Page	What it covers
Harness Engineering	Agent = Model + Harness, feedforward vs feedback, computational vs inferential controls, humans on the loop (Böckeler/Fowler)
Guides	feedforward controls: templates, examples, writing style, conventions
Sensors	deterministic feedback: structure checks, the sensor runner, hard gates
Evals	inferential feedback: LLM-judge rubrics, threshold 8.0, retry, spec-satisfied
Pipeline and Stages	the six gated stages, anatomy of a stage, markers, tokens, status bar, SDD loop
Golden Path	the paved road, the five properties, detours, per-discipline paving
Agents	product-manager, staff-software-engineer, system-architect
Designer Skill	M3, dark/light theme, modern type, Behance polish, i18n, favicon for new UIs

Start here: Harness Engineering explains the whole mental model; everything else is a detailed view of one part.

v5: autonomy and subagents

The pipeline is moving from humans in the loop (answer an input before every artifact) to humans on the loop (approve direction at two gates, everything between runs on its own), carried by an orchestrator that dispatches small, single-purpose subagents.

Page	What it covers
Autonomy	the intake stage, resolve-mark-proceed, gated autonomy, feedforward over feedback
Orchestration and Subagents	orchestrator + leaf subagents, the adversarial writer/critic split, model tiering, Claude Code constraints

These pages describe the v5 model, now live across every stage. The canonical shape is .claude/shared/pipeline-pattern.md; a couple of refinements (Haiku tiering, per-org repos.md) are still marked (planned) inline.

System design

The system-architect agent and its topic playbooks. Each episode of the System Design series becomes a skill plus a detailed page here: theory, reference architecture, trade-offs, and citations.

Page	Design problem
System Design Method	the 13-stage method, the three pillars, back-of-envelope numbers, the canon
URL Shortener (#1)	read-heavy key lookup, short-code generation, cache, abuse
Rate Limiter (#2)	distributed throttling, token bucket, fail-open, multi-region budgets
Search Engine (#3)	web crawler, indexing, ranking, query serving at scale

More episodes land as new playbooks and new pages. Candidates: API gateway, distributed queue, news feed, distributed cache, notification system.

How the two halves connect

A System Design Doc is an optional front stage before the main pipeline: a strong design feeds a sharper PRP and a grounded plan. The same harness shape, guides, sensors, evals, gated approval, runs under both halves.

Global references

• Birgitta Böckeler, Harness engineering for coding agent users & Maintainability sensors for coding agents, martinfowler.com, 2026.
• Martin Kleppmann, Designing Data-Intensive Applications, O'Reilly, 2017.
• Jeff Dean, Designs, Lessons and Advice from Building Large Distributed Systems, LADIS, 2009.
• Werner Vogels (design for failure), Pat Helland (immutability, distributed transactions), Michael Nygard (Release It!), John Ousterhout (A Philosophy of Software Design).