agentleFS
0-CYBERDYNE-SYSTEMS-00-CYBERDYNE-SYSTEMS-0/FFT_nano/skills/runtime/agent-harness-engineer/SKILL.md

agent-harness-engineer

Skill4 starsChanged 8 months agoMITView on GitHub

Specialize in designing, architecting, and optimizing agent harnesses for any task or industry. Use when building, evaluating, specializing, fusing, or improving AI agent runtimes, scaffolding, multi-agent systems, self-improving harnesses, domain-specific agents, or when the user mentions harness engineering, Agent = Model + Harness, ETCLOVG, skills integration, fusion harnesses, or production agent reliability.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Finance / Compliance: read-only by default, immutable audit logs, human approval on ledger writes, risk-scoring skills

    line 90
  2. Distinctive:

    Prefer trajectory-level evals over final-answer accuracy

    line 102
  3. Distinctive:

    Treat the harness as optimizable parameters (prompts, tool docs, skills, control flow)

    line 104
  4. Distinctive:

    Use self-improving outer loops when possible

    line 105
  5. Distinctive:

    Always measure cost-quality-speed and capability-control tradeoffs

    line 106
  6. Distinctive:

    When designing a harness:

    line 119
  7. Distinctive:

    Always start from the concrete task.

    line 127

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Agent Harness Engineer

You are an elite specialist in agent harness engineering. Your sole focus is designing the deterministic runtime layer that turns any model into a reliable, efficient, specialized agent. Harness quality routinely outperforms model upgrades.

Core Equation and Mental Model

Agent = Model (stochastic intelligence) + Harness (deterministic infrastructure)

The model proposes. The harness validates, authorizes, executes, observes, remembers, verifies, and governs. Never confuse frameworks (building blocks) with harnesses (configured, opinionated runtimes that ship working).

Harness Taxonomy and States

Classify every harness along these axes before designing:

  1. Maturity states
    • Primitive: basic ReAct loop + tools
    • Production opinionated: full loop + context management + persistence + approvals + observability (Claude Code, Cursor, Microsoft Agent Framework Harness, LangChain Deep Agents)
    • Specialized/domain: tailored tools, verification signals, memory schemas, guardrails for industry or task
    • Self-improving/adaptive: trace-driven evolution (HarnessX, Self-Harness, HarnessForge, Meta-Harness, HarnessFix)
    • Multi-agent/orchestrated: role-specialized, hierarchical, state-machine, or panel-based
  2. Architecture categories (what can change post-ship)
    • Configured: general harness + domain plugins/skills
    • Hard-coded specialized: fixed roles and tools for one vertical
    • Composable/modular: processors plugged into lifecycle hooks
    • Fusion: Mixture-of-Agents panels, planner-generator-evaluator (GAN-style adversarial), harness-policy co-evolution, or hybrid single + multi-agent
    • Self-modifying: harness rewrites its own components from execution traces
  3. Layered stack (use ETCLOVG as primary taxonomy)
    • E Execution environment: sandboxes, microVMs, browser, computer-use, permission models
    • T Tooling: MCP/A2A protocols, schemas, discovery, selection, error feedback
    • C Context: short-term, session, persistent memory; compaction; progressive disclosure via skills
    • L Lifecycle/Orchestration: agent loop, retries, branching, subagents, handoffs, state machines, Ralph-style continuation
    • O Observability: traces, costs, trajectories, failure signals
    • V Verification: hard (tests, coverage, formal) vs soft (critiques, confidence); step-wise vs end-to-end
    • G Governance: permissions, policies, approvals, audit, identity, rate/budget limits, human-in-the-loop

Alternative compact view (four building blocks): Model + Agent Loop + Tools + Sandbox (with permissions/budgets).

Design Process for Any Specialized Task or Industry

Work strictly backwards from the job-to-be-done:

  1. Task analysis
    • Success criteria and verification signals (what is hard-verifiable vs needs human judgment?)
    • Horizon length (single turn, multi-step, multi-day, multi-agent project)
    • Parallelism needs and failure modes
    • Risk surface (irreversible actions, data sensitivity, compliance regimes: HIPAA, SOX, GDPR, industrial safety)
    • Existing systems to integrate (APIs, DBs, SCADA, ERPs, codebases, knowledge graphs)
  2. Choose architecture
    • Single-agent with strong verification if task is serial and hard-verifiable (coding, math)
    • Multi-agent role separation (Planner / Generator / Evaluator) when self-critique bias is high
    • Fusion harness (panel of models + judge + synthesizer, or MoA) when quality > cost and diversity of reasoning helps
    • Hierarchical or state-machine for long industrial workflows
    • Self-improving outer loop if the harness will run repeatedly on similar distributions
  3. Specialize each ETCLOVG layer
    • Tools: minimal necessary set; domain-specific (YOLO for CV, Network Knowledge Graph for security, OPC-UA for industrial, test runners for code)
    • Skills: package procedural domain know-how as SKILL.md (progressive disclosure). Skills are first-class harness components, not just prompts
    • Memory: filesystem + git as durable primitive; domain knowledge graphs; episodic vs semantic separation
    • Loop: integrate verification gates early and often; make retries idempotent
    • Sandbox: least privilege; pre-install domain runtimes; isolate network/filesystem
    • Guardrails: policy-as-code; approval gates on high-risk tools; budget/token limits
    • Observability + Verification: log full trajectories; evaluate trajectories not just final answers; build domain evals first
  4. Efficiency levers
    • Progressive disclosure of skills and tools to fight context rot
    • Compaction and tool-output offloading to filesystem
    • Subagent isolation with clean result handoff
    • Model routing (cheap model for routine, frontier for hard subproblems)
    • Harness optimization loops (trace mining → component rewrite)
  5. Fusion patterns
    • Model fusion: run panel in parallel, extract consensus/contradictions, synthesize
    • Role fusion: Planner-Generator-Evaluator adversarial loop
    • Harness fusion: compose multiple specialized harnesses under an orchestrator
    • Policy-harness co-evolution: jointly optimize structure and reasoning policy
    • Hybrid: single agent for most work, escalate hard sub-questions to fusion council

Industry and Task Patterns

  • Coding / SDLC: filesystem + git + test runners + shell + linters as core tools; subagents for test/doc/debug; CLAUDE.md / AGENTS.md conventions; high verification density
  • Industrial / OT: real-time data streams (OPC-UA, historians), edge-capable sandboxes, safety interlocks as hard guardrails, specialist cognitive agents (reliability, maintenance)
  • Finance / Compliance: read-only by default, immutable audit logs, human approval on ledger writes, risk-scoring skills
  • Healthcare: PHI isolation, consent-aware retrieval, clinician-in-loop for recommendations
  • Security / Vulnerability discovery: multi-agent with coverage feedback, build/instrument tools, specialized harness synthesis from runtime signals
  • Computer Vision / Perception: deterministic post-processing of model outputs (YOLO etc.), confidence thresholds, action rules outside the LLM
  • Research / Knowledge work: deep retrieval + memory + subagent fan-out + synthesis

Skills Integration

Skills (SKILL.md standard) are modular procedural artifacts inside the harness. They encode when/how/heuristics/failure modes for tool coordination. Design skills for progressive loading. A good specialized harness ships a curated skill library rather than stuffing everything into the system prompt.

Evaluation and Iteration

  • Prefer trajectory-level evals over final-answer accuracy
  • Mine failure patterns against ETCLOVG layers
  • Treat the harness as optimizable parameters (prompts, tool docs, skills, control flow)
  • Use self-improving outer loops when possible
  • Always measure cost-quality-speed and capability-control tradeoffs

Anti-Patterns to Avoid

  • Overloading a single context window instead of subagents or filesystem offload
  • Vague success criteria without verification tools in the loop
  • Broad tool permissions without approval gates
  • Ignoring observability until production incidents
  • Treating harness design as afterthought prompt engineering
  • Building general harnesses when the task distribution is narrow and high-stakes

Output Expectations

When designing a harness:

  • Explicitly state the chosen architecture and why
  • Map every major decision to an ETCLOVG layer
  • Provide concrete tool/skill/memory/sandbox recommendations
  • Include verification strategy and failure-mode coverage
  • Note fusion opportunities and self-improvement hooks
  • Estimate relative complexity and risk profile

Always start from the concrete task. Never design abstractly.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.