#agent-harness

A clean-room Go implementation of agentic coding harness patterns, derived from analyzing production AI agent architectures. Built for learning, extending, and teaching how to build coding agents like Claude Code, OpenCode, and Gemini CLI. Supports OpenRouter and Anthropic out of the box.

Enforces plan → code → tests → validate → docs discipline for AI coding agents (Claude Code, OpenCode).

The model-agnostic execution layer for AI agents: scoped routing and delegation, isolated worktrees, evidence, observability, and learning loops. The stack, not another wrapper. Seven executors — Claude Code, Codex, Cursor, opencode, DeepSeek Harness, Hermes, Pi — under one metered contract.

A local-first multi-agent coding harness that connects model providers and ACP agents, binds them to specialized roles, and coordinates them in a shared session.

Keep Claude Code and Codex aligned with task requirements through intent proof, drift checks, visible run state, and evidence-based verification.

Executable engineering judgment for AI coding tools: blueprints, gates, checkpoints, policy, memory, and audit trails.