best-of-Agent-Harnesses Now Ranks 141 Harnesses Across 12 Categories — and Ships an MCP Server So Agents Can Pick Their Own Runtime
RyanAlberts/best-of-Agent-Harnesses (452 stars, last rescored August 2, 2026) catalogues 141 agent harnesses — the runtime layer that turns model reasoning into sustained tool-using, error-recovering behavior — across 12 categories including progressive disclosure, coding agents, multi-agent orchestration, memory layers and evaluation. Beyond star ranking it applies four editorial dimensions: a four-tier simplicity↔capability scale, headless-ready (★, designed for unattended runs), durable execution (✱, state survives restarts), and open-source signal. The machine-readable surface is the real point — `harnesses.json`, `llms.txt`, and an official MCP server (`io.github.RyanAlberts/agent-harnesses`) exposing `recommend`, `compare`, `pick_harness` and `search_harnesses` so an agent can select its own harness.
Source
↳ Follow the thread