Harness Development | ECOSYSTEM Research

GOOSE RUST BASED AGENT https://goose-docs.ai/ https://github.com/aaif-goose/goose Open Hands https://www.openhands.dev/ https://github.com/OpenHands/OpenHands Aider (Python) https://github.com/Aider-AI/aider - Smars-Bin-Hu/copilot-agentic-harness: An agentic framework that wraps Copilot to add task persistence across chats, hook-enforced guardrails, human-approved write-backs, and observability. This is the closest you'll get to a "Copilot harness" repository, but it's an orchestration layer, not the core engine. - davidreis97/defending-code-reference-harness-copilot: A port of Anthropic's Claude Code reference harness, adapted to run on GitHub Copilot for autonomous vulnerability discovery. Studying this is valuable because it shows how to build a security-focused coding harness whose backend happens to be Copilot. - robmab/one-agent-harness: A lean harness template (6 skills, 1 sensor, 1 agent) that works with both Claude Code and GitHub Copilot. If your goal is to learn how a production coding agent is architected, these are not the primary sources—they are integration layers. For that, the Rust harnesses below are far more instructive. 🦀 Rust Coding Harnesses (Coding Specialists, Not General Assistants) The Rust ecosystem has a growing number of coding-specific agent harnesses. They fall into a few clear categories, and several are direct ports of well-known TypeScript agents, which means you can compare implementations side-by-side. Full-Featured Terminal Agents (Original Rust Implementations) Tau (nwyin/tau) A Rust coding-agent harness with a full-screen TUI, headless CLI, and JSON-RPC serve mode. Its architecture is explicitly layered: ai (provider adapters, streaming), agent (generic runtime, context compaction), and coding-agent (the built-in coding harness with tools, permissions, sessions, and traces). It includes a built-in tool suite for file edits, shell, search, and planning, plus a persistent Python REPL tool and skill discovery from .tau/skills/ directories. This is a strong reference for a clean separation between generic agent infrastructure and coding-specific logic. Rho (matthewyjiang/rho) A lightweight Rust agent harness with a small footprint and opinionated defaults. It is fast and small, supports a wide range of providers (including Copilot and local Ollama), and ships an embeddable SDK for building headless Rust agents with explicit providers, tools, sessions, and cancellation. Rho also includes coding tools (read, edit, search, shell, web) out of the box. For a low-level developer, Rho is interesting because it is designed to be embedded and composed, not just run as a monolithic CLI. VTCode (vinhnx/VTCode) A Rust coding agent built for long-running autonomous workflows, with OS-native sandboxing, multi-provider support, and open protocols. Its feature set is unusually deep for a Rust harness: ripgrep search, ast-grep symbol maps, worktree isolation for parallel agents, propose/verify sub-agent separation, durable loop state, and cost guardrails. The worktree isolation and propose/verify separation are directly relevant to building a robust coding harness that can recover from failures. CodeWhale (jcode / CodeWhale project) A terminal coding-agent harness that has gained significant traction (40k+ stars as of mid-2026). It treats governance as a runtime concern: identity, user intent, approval gates, sandboxing, rollback, and recoverable sessions are all part of the agent runtime rather than afterthoughts. It uses a constitution.json for persistent constraints and supports side-git snapshots for rollback without polluting the project repository. If you care about safety and recoverability in a coding agent, CodeWhale's design is worth studying. Ports of Existing Coding Harnesses (Rust Rewrites) These are especially useful for learning because you can compare the Rust implementation against the original TypeScript or Python source. Pi (nktkt/pi, c4pt0r/pie) Pi is a coding-agent harness focused on the core coding-agent loop. The Rust port (nktkt/pi) is a workspace with pi-ai (streaming provider APIs), pi-agent (the agent loop with per-tool permission gates and typed errors), and pi-coding-agent (the CLI with session persistence and an AGENTS.md/CLAUDE.md loader). The crates are published on crates.io, so you can reuse the runtime directly in your own harness. c4pt0r/pie is another Rust rewrite of the same pi project, focused on a terminal-based coding agent that can inspect files, make edits, run shell commands, and continue previous sessions. Tack (Tack-AI/tack) A Rust reimplementation of the pi coding agent that goes beyond a port: it adds LSP navigation and diagnostics, file checkpoints, persistent memory, sub-agents with worktree isolation, cross-session search, cron, OS sandboxing, declarative permissions, and an MCP client/server. It is wire- and storage-compatible with the TypeScript pi and supports 42 built-in providers (including local Ollama and llama.cpp) with an embedd…