A curated, verified, cross-vendor directory of AI agent harnesses — the runtime scaffolding that turns a model into an agent.
◆ 74 harnesses profiled · 9-axis capability schema · last verified 2026-07-08
The clean-room Rust/Python rewrite of Claude Code's harness architecture born hours after the March 31, 2026 source leak — reportedly the fastest-growing repo in GitHub history (~195k stars) with permission-gated tools, MCP, agent swarms, and container-first deployment — though the project is chaotic, largely agent-maintained, and its details vary across sources.
Mario Zechner's radically minimal MIT/TypeScript harness (~68k stars): four tools (read/write/edit/bash), a ~300-word system prompt, deliberately no built-in MCP/subagents/permission-popups, with everything else added via TypeScript extensions, skills, and npm/git-shareable packages across interactive, print/JSON, RPC, and SDK modes.
Block's Rust-built, Apache-2.0 general engineering agent — now stewarded by the Linux Foundation's Agentic AI Foundation — distinguished by the deepest MCP integration in the ecosystem (70+ extensions, can also act as an MCP server) and shareable YAML 'Recipes' that scaled agent workflows to 60% of Block.
The original terminal AI pair programmer (since 2023): git-native supervised editing with auto-commits per change, a tree-sitter repo map, architect/editor model pairing, and watch-mode AI comments — still largely a one-maintainer community project with releases roughly every two weeks.
The NeurIPS-2024 research harness from the SWE-bench team that pioneered the agent-computer-interface concept: takes a GitHub issue and autonomously fixes it in a containerized environment with any LM, and remains the reference batch-evaluation harness for academic agent research.
A batteries-included 2026 fork of Pi (created Dec 31, 2025; ~17k stars) that inverts Pi's minimalism: hash-anchored edits that reject stale patches, LSP and DAP debugger integration, in-process ripgrep/bash, 40+ providers, 32 tools, and subagent orchestration.
Terminal agent purpose-built for very large multi-file tasks — 2M-token effective context via tree-sitter project maps, a cumulative diff-review sandbox that stages changes outside the repo, and configurable full-auto mode — but Plandex Cloud wound down in Oct 2025 and the repo has had no pushes since, leaving it self-host-only and dormant.
A deliberately radical ~100-line Python agent with bash as its only tool (no tool-calling API, no config sprawl) that still scores >74% on SWE-bench Verified, used by Meta, NVIDIA, and IBM as a baseline/benchmarking harness and proof that scaffolding can be minimal.