harness-design

---

name: harness-design description: Design and set up an agent harness for any project — first judge the verification signal (can success be machine-checked? are failures reversible? can the work be split?), then choose a structure: solo agent + check command, pipeline, Planner → Builder → QA triangle, or fan-out workers. The triangle is one structure, not the default. Use this skill whenever the user wants to build with multiple AI agents, set up automated workflows, create an agent pipeline, build an eval harness or an experiment/optimization loop, apply harness methodology, or structure any complex task that benefits from separating planning, execution, and verification. Also use when the user mentions "harness", "multi-agent", "planner/builder/qa", "coordinator/compiler", "generator/evaluator", or wants to coordinate multiple Claude Code instances.

Harness Design

A meta-skill for setting up multi-agent harnesses. The output is the scaffolding (agents, blackboard, hooks) that the project uses to do its real work — not the work itself.

This skill is evidence-based: it updates after each real project from observations recorded in evidence/. See CHANGELOG.md for version history and evidence/ for the raw findings.

---

Three principles (the only invariants)

1. Plan ⊥ Execute — the agent that converts user intent into a markdown plan must be a different context from the agent that executes the plan.
2. Execute on documents — every executor reads from a markdown blackboard and writes a markdown report. No verbal handoffs, no implicit state.
3. Generate ⊥ Evaluate — the agent that produces work cannot be the agent that judges it. The evaluator runs the product, not just reads the code. When an oracle closes the verification loop, the evaluator can be a command, not an agent.

These invariants govern multi-agent structures. Step 1 may route a task out of the framework entirely (solo + check), where the check command carries the Evaluate role.

The hidden fourth — Context Economy

Role separation is not aesthetic and not about "self-persuasion bias." It is context sharding:

• The Coordinator's context must stay lean across the project's lifetime to make sound long-horizon decisions.
• Generators and Evaluators are sacrificial contexts — they fill up doing dirty work and die at task end.
• The blackboard (.harness/ files) is how state survives the death of sacrificial contexts.

This reframe matters: it tells you when role separation can be relaxed (a trivial Coordinator-side fix is cheaper than a Builder spawn) and when it must hold (anything that would force Coordinator to Read multiple source files).

See references/context_economy.md.

---

Step 0 — Detect mode

Does the target project already have .claude/agents/ with agent definitions?

• NO → New build. Go to Step 1.
• YES → Upgrade. Read references/upgrade_playbook.md.

Step 1 — Choose the structure (before any scaffolding)

Question 0 — prequalifier. Is there any signal — machine-checkable or human-judgeable — that consistently tells success from failure? If the work is pure product judgment, vague intent, or taste calls with no anchor: do not enter the orchestration framework. Stay human-led until a checkable signal exists.

Then three questions, in order:

1. Oracle — can success/failure be judged by a command (tests, compiler, exit code, objective metric, fixture diff)? Or is it partly subjective (UX, tone, intent alignment, citation quality)?
2. Blast radius — is a wrong action reversible (sandbox, git, rerunnable), or expensive/irreversible (production data, real deploys, money)?
3. Splittability — one tightly-coupled body of work, a fixed sequence of stages, or many independent sub-tasks?

Route on the answers — one structure, not always the triangle:

Answers	Structure
Oracle-closed + reversible + single body	Solo agent + check command — the check command is the evaluator; no QA agent
Oracle-closed + reversible + ordered stages	Pipeline — each stage gate is a command; no resident QA agent
Subjective verdicts needed, or high blast radius (even with an oracle)	Planner → Builder → QA triangle — an independent evaluator context earns its cost here
Subjective + many independent units (exploration, research, eval batches)	Fan-out workers + judge — N generators ⊥ an independent judge

Too small to harness: if the whole task fits in one session and a check command already exists — don't scaffold. Do the task, run the check.

Unsure, or hybrid (e.g. a triangle with a fan-out stage inside)? Read references/structure_routing.md for worked examples and calibration recipes.

Step 2 — Instantiate the chosen structure

If routing landed on the triangle — three roles, one blackboard, two .claude/ files:

.claude/
├── agents/
│   ├── coordinator.md
│   ├── builder.md
│   └── qa.md
└── settings.json   ← "agent": "coordinator" + Stop hook QA gate

.harness/
├── spec.md         ← current plan (Coordinator-written, user-confirmed)
├── progress.tsv    ← one row per QA round
├── HANDOFF.md      ← cross-session continuity
├── lessons.md      ← rolling failure → fix → prevention (v0.1)
├── CONTEXT.md      ← domain glossary, ubiquitous language (v0.2, when /grill-with-docs is used)
├── docs/adr/       ← architectural decision records (v0.2, three-condition rule — see references/software_harness_with_skills.md)
└── reports/        ← build_*.md + qa_*.md per round

For the other structures the minimal file sets are smaller — solo: GOAL.md + check.sh + LOG.md; pipeline: PIPELINE.md + per-stage checks; fan-out: SCOPE.md + run.sh + judge + findings/. See references/structure_routing.md.

Routing table — read only what you need:

If you need …	Read
Choosing between solo / pipeline / triangle / fan-out (worked examples, calibration recipes)	references/structure_routing.md
Software project + skill packs (mattpocock methodology layer; gstack runtime retired 2026-06)	references/software_harness_with_skills.md
Software project blackboard, agent files, browser verification	references/software_harness.md
Knowledge / wiki / research harness (wiki-as-blackboard)	references/knowledge_harness.md
Operations / continuous-loop harness (experience layer)	references/operations_harness.md
YAML frontmatter, tool allocation, MCP integration	references/agent_definitions.md
Hooks, settings.json, what to enforce vs warn	references/enforcement.md
Rolling lessons capture + maintenance	references/lessons_pattern.md
Why role separation = context economy + practical rules	references/context_economy.md
Upgrading an existing harness without breaking it	references/upgrade_playbook.md

Step 3 — Do NOT add by default

Earlier versions of this skill carried extra scaffolding that the finsim audit (evidence/2026-04-29_finsim_audit.md) showed did not earn its keep. Skip unless project data justifies it:

• test.md acceptance file (finsim never wrote one; nothing failed)
• Duration / domain ritual classification table (judgment is faster)
• 15-item instantiation checklist (compressed to 5 below)
• "Sources read-only" invariant (only applies when there's an input data layer)
• Long anti-pattern enumerations in the main skill (each one belongs in a real lesson, dated and commit-linked)
• 2/2 consecutive-PASS dynamic exit rule (real trigger rate: 3% in finsim — vestigial)

Use the simpler rule: r1 PASS exits the unit. r1 FAIL → r2. Same fail again → re-plan and write a lesson. This is the canonical exit rule — the reference files point here.

Step 4 — Instantiation checklist (5 items)

• Role files for the chosen structure — triangle: coordinator / builder / qa in .claude/agents/; solo + check: none (just GOAL.md + the check command); pipeline / fan-out: see references/structure_routing.md
• Blackboard for the chosen structure — triangle: .harness/ with spec.md, progress.tsv header, lessons.md header; solo: GOAL.md + LOG.md
• Gate wired in .claude/settings.json — triangle / fan-out: "agent": "coordinator" + agent-type Stop hook QA gate; solo / pipeline: Stop hook runs the check command
• CLAUDE.md describes project rules + harness roles in ≤ 2 screens
• First task spec written and user-confirmed before any code is touched

Anti-patterns observed in the wild

Only four, all evidence-backed (see evidence/):

1. Coordinator dives in to "fix it quick" repeatedly → context bloat, eventual compaction loss of early decisions. The fix it produced is fine; the long-tail context degradation is invisible. → references/context_economy.md.
2. Same failure happens twice without a lesson written → known gotcha keeps biting. finsim's PR-CALENDAR-1 r1 FAIL hit a gotcha CLAUDE.md explicitly warned about, because nothing forced new gotchas back into the static rules. → references/lessons_pattern.md + PostToolUse hook in references/enforcement.md.
3. QA reads the diff but doesn't run the product → ships visual / runtime bugs that look fine in code review. finsim's PR-AUTH-1 stageC r1 FAIL (inline style overriding CSS height) was only caught because QA actually rendered the page. → references/software_harness.md (browser verification patterns).
4. Spec becomes a wishlist nobody verifies against → scope drifts silently. Build report should cite spec line numbers; QA verdict references both.

Evidence-based iteration

This skill updates as evidence arrives:

• evidence/{date}_{project}.md — dated findings from past projects (audits, post-mortems, what worked, what didn't).
• After each real project ships, write a short evidence entry.
• Promote stable findings into SKILL.md or the references; demote outdated guidance to historical notes (don't delete — keep the trail).
• CHANGELOG.md records what changed and which evidence drove it.

See evidence/README.md for the template.