harness-creator

<p align="center"> </p>

<p align="center"> Design safer, sharper, production-ready agent harnesses instead of vague "AI agent" architectures. </p>

harness-creator is a Codex skill for analyzing, designing, and implementing production-ready agent harnesses.

It helps you turn a fuzzy request like:

• "Build a coding agent"
• "Make this production-grade"
• "Add subagents and planning"
• "Design an MCP-enabled agent runtime"

into a concrete, layered runtime design with clear module boundaries, implementation order, and safety rules.

Most agent designs fail in one of two ways:

1. They stay too abstract and never become a buildable system.
2. They jump straight to orchestration and autonomy before the core loop, tool safety, persistence, and verification are dependable.

harness-creator fixes that by forcing the conversation toward:

• a real loop/tool/permissions/memory/orchestration model
• a progressive build order
• clear distinctions between facts, recommendations, and unknowns
• explicit safety and verification requirements

• A progressive architecture ladder for agent runtimes.
• A concrete implementation order from minimal loop to advanced orchestration.
• A required output contract so analysis does not stop at "I inspected these files."
• Strong guidance for permissions, planning, subagents, memory, compaction, integrations, and verification.

• Loop first, orchestration last: prevents premature swarm complexity.
• Deny-first safety: pushes shell/file/network safety into the harness, not just prompts.
• Facts vs inference: helps you avoid overclaiming when incomplete information exists.
• Implementation-ready: produces module boundaries and execution order, not just principles.

This skill was pressure-tested in practical Codex workflow scenarios.

Ob