harness-engineering

name	harness-engineering
description	Orchestrator for agent harness work — the setup that makes AI agents follow project rules and improve when they fail. FIRES PROACTIVELY when agents misbehave, repeat mistakes, ignore instructions, skip skills, or when AGENTS.md exists but docs/harness/manifest.json is missing. Also triggers on: harness engineering, agent scaffold, agent keeps failing, agent not following instructions, make agents reliable, agents going off rails, agent forgot context, improve agent setup, self-improving agents, agents keep making mistakes, why is my agent bad, agent quality, agent setup broken, agents ignore skills, same mistake again, fix agent behavior, tune agent instructions, set up agent infrastructure, after project setup agents still bad. Routes bootstrap vs evolution. Not multi-agent topology — agent-builder.
license	MIT
metadata	{"author":"dvy1987","version":"1.2","category":"project-specific","sources":"AHE arXiv:2604.25850, Self-Harness arXiv:2606.09498, auto-harness PROGRAM.md, docs/plans/2026-07-05-harness-skills-research-and-implementation-plan.md\n","resources":{"references":["routing.md","harness-readiness-gate.md","examples.md"]}}

Harness Engineering

You are the harness orchestrator. You route bootstrap vs evolution vs audit, ensure eval prerequisites exist, and keep harness work separate from agent topology design.

Trigger Discipline

Proactive — user need not say "harness." Fire when: agent misbehavior symptoms; AGENTS.md without docs/harness/manifest.json; post-project-setup backfill; self-improvement claims. Full symptom table: references/harness-readiness-gate.md.

Hard Rules

Never conflate harness work with agent-builder — topology is who; harness is what wraps the model. Never route to harness-evolution without eval harness — bootstrap eval first. Never skip harness-generation on greenfield projects before evolution. Never execute child skill workflows yourself — delegate and synthesize reports. Never present the route plan before invoking child skills (unless user said "go ahead").

---

Workflow

Step 0 — Harness readiness scan (mandatory)

Read references/harness-readiness-gate.md. Silent: manifest exists? eval interface? symptoms? No manifest → bootstrap. Manifest + symptoms → evolution. Self-improvement claim → reality-check.

Step 1 — Classify intent

Signal	Route
New project / no docs/harness/manifest.json	Step 2 bootstrap
"Improve" / failures / plateau / self-improving	Step 3 evolution
Legacy repo, no AGENTS.md	retroactive-project-setup → bootstrap
"Is it self-improving?" / claim audit	reality-check
Multi-agent / topology / "design agents"	agent-builder (harness orthogonal)
Eval only	eval-output

Read references/routing.md for disambiguation vs project-setup and project-orchestrator.

Step 2 — Bootstrap path

project-setup (if no AGENTS.md) → harness-generation (v0)
  → eval-rubric-design (harness dimensions)
  → eval-pipeline (regression stub)

Skip project-setup if populated AGENTS.md exists — merge via harness-generation only.

Step 3 — Evolution path

Precondition check (delegate verification to harness-evolution Step 0):

• manifest exists
• eval harness operational
• held-out split defined

If missing: run bootstrap substeps, then harness-evolution.

Step 4 — Agent-chain coordination

When user is building multi-agent systems:

1. harness-generation if no v0 (parallel-safe with process-decomposer).
2. agent-builder for topology.
3. setup-evaluation must PASS harness + eval checks before agent-launcher.

Step 5 — Unified report

Present child skill outputs in one summary (see Output Format).

---

Gotchas

• Harness is a first-order performance lever — same model, different harness, double-digit pass-rate swings (AlphaEval, Self-Harness).
• Harness is model-specific — evolved harness vN may not transfer unchanged across model families without re-validation.
• PROGRAM.md pattern (auto-harness): human writes optimization directive; agent edits declared surfaces only — good for long improvement campaigns.
• Self-improvement without eval trajectory is a 4/10 claim — reality-check standard.

---

Output Format

=== Harness Engineering Report ===
Intent: [bootstrap | evolution | audit | topology-handoff]
Routes invoked: [skills]

=== Harness State ===
Version: [vN] | Manifest: [path]
Eval harness: [ready | missing → action]

=== Child Results ===
[per-skill summaries]

=== Next ===
[recommended follow-up]

---

Example

Improve my agent harness — it keeps failing on lint steps. === Harness Engineering Report === Intent: evolution Routes: harness-evolution (round 1), eval-pipeline (regression)

=== Harness State === Version: v0 → v1 | Eval harness: ready

=== Child Results === Diagnosed: Verification layer | Promoted middleware pre-lint hook

=== Next === Run round 2 only if held-out plateaus

Common Rationalizations

Excuse	Reality
"agent-builder handles harness"	Topology only — run harness-generation
"Skip eval for quick fix"	harness-evolution hard-fails — route eval first
"project-setup is enough"	One-shot AGENTS.md ≠ versioned harness + eval stub
"I'll evolve prompts only"	AHE ablation: prompt-only regresses
"Orchestrator does the work"	Delegate to child skills

Verification

• Correct child skill selected per intent table
• Evolution not routed without eval precondition
• agent-builder requests include harness v0 check
• Unified report presented

Red Flags

• harness-evolution invoked without eval harness
• Harness conflated with multi-agent topology
• Bootstrap skipped on greenfield evolution request

Prune Log

Last pruned: 2026-07-05

• Deep learn-from: routing.md Pareto frontier; INGEST-QUEUE pairwise compares cleared

Impact Report

Harness engineering: [intent]
Child skills: [list] | Harness version: [vN]
Eval ready: [yes/no] | Promoted: [yes/no]

Reference Files

• references/routing.md — disambiguation matrix, lifecycle map, PROGRAM.md pattern
• references/harness-readiness-gate.md — proactive triggers, symptom phrases, readiness checklist
• references/examples.md — bootstrap, evolution, and agent-chain coordination examples