best-of-Agent-Harnesses

The best AI agent harnesses in 2026, ranked by category

These are the top three agent harnesses in each of the 12 categories of best-of-Agent-Harnesses, a hand-curated list of 167 harnesses re-ranked from live GitHub data every week; the stars below were captured 2026-09-20, and the page regenerates with every refresh.

How this page is ranked

Projects enter the list by editorial review against one rubric: does it own the agent loop, the tool wiring, the approval model, or the memory, and is it maintained. Within a category the order is GitHub stars, captured weekly, and each project carries its adoption-surface tier and its autonomy and recovery scores, so a high star count never hides a harness that cannot run unattended or survive a crash. Archived projects move to the graveyard and never appear here. The six-question guide turns these tables into a decision, and why the harness matters more than the model explains why the choice deserves the care.

Progressive disclosure harnesses

Formats, runtimes, and patterns that reveal context, tools, or instructions in layers—index first, details on demand—to control tokens and improve agent focus (the "map, not encyclopedia" principle). Full category →

  1. Headroom (73.2k stars, mostly simple, not a loop-owning harness (a format, config, or library)): Compresses tool outputs, logs, files, and RAG chunks with content-aware compressors before they reach the model—claimed 20% fewer tokens for coding agents and 60–95% fewer for JSON, same answers. Ships as a library, HTTP proxy, or MCP server, so it drops in front of whatever harness you already run.
  2. awesome-cursorrules (40.8k stars, super simple, not a loop-owning harness (a format, config, or library)): Curated .cursorrules and skills that leverage Cursor's index-then-load model; the canonical collection for rules-as-progressive-disclosure in the IDE.
  3. agents.md (24.5k stars, super simple, not a loop-owning harness (a format, config, or library)): Open format for repo-scoped agent briefings; nested AGENTS.md files scope instructions per directory, so agents get a map of what exists and load only what's relevant. Read by 20+ tools including Codex, Cursor, and Copilot.

Coding agent products (IDEs, CLIs, full suites)

Turnkey coding agents you install and run: IDE extensions, terminal CLIs, Dockerized workspaces. Each entry notes which part is the harness (the agent loop, tool wiring, approval model) versus the UI shell (VS Code extension, TUI, browser client). Full category →

  1. opencode (209k stars, slightly complex, autonomy headless, recovery resumable): Open-source terminal coding agent (formerly sst/opencode; transferred to anomalyco). The harness is a multi-provider tool-call loop (Claude, OpenAI, Gemini, local) with strong plugin and MCP support; the TUI is the shell. 100% OSS, very actively shipped.
  2. Codex (125k stars, slightly complex, autonomy bounded, recovery resumable): OpenAI's terminal coding agent. The harness is the sandboxed tool-call loop with multi-provider support; the CLI is the shell. Reference implementation for "official CLI that ships code."
  3. pi (108k stars, slightly complex, autonomy bounded, recovery resumable): The upstream AI agent toolkit behind this list's oh-my-pi fork: a unified multi-provider LLM API, agent loop, and TUI shell providing the harness that oh-my-pi's Rust rewrite builds on.

Coding harness configs and SDKs

Skill packs, slash-command libraries, meta-prompting frameworks, and official SDKs that give you the harness (the agent loop, planning, memory, hooks) without bundling a specific IDE or CLI shell. Full category →

  1. superpowers (289k stars, complex, not a loop-owning harness (a format, config, or library)): Performance-oriented harness pack for Claude Code and 13 other harnesses (Codex, Cursor, OpenCode, Gemini CLI, more): skills, instincts, memory, security, research-first workflows. Treats harness engineering itself as the performance lever.
  2. Anthropic Skills (177k stars, mostly simple, not a loop-owning harness (a format, config, or library)): Anthropic's official Agent Skills repository: SKILL.md-based folders (instructions, scripts, resources) Claude dynamically loads on Claude Code, Claude.ai, and the API. The reference for progressive-disclosure skill packs in 2026.
  3. GStack (134k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Garry Tan's Claude Code skill stack: 23 slash-command modes (CEO/eng/design review, QA, ship, browse, retro, …) that structure one assistant as a virtual engineering team. Daily driver while running YC.

Personal agent runtimes

Always-on, self-hosted agents you run as a daemon and talk to from chat apps: gateway runtimes, second brains, and self-improving assistants. The agent as a product you operate, not a library you build with. Full category →

  1. OpenClaw (390k stars, complex, autonomy headless, recovery resumable): Self-hosted, always-on personal agent (formerly Clawdbot/Moltbot): a gateway + event-loop runtime that treats messages, heartbeats, crons, and webhooks as one input queue, persists state to local files, and lives in your chat apps (WhatsApp, Telegram, Slack, Discord). 13,700+ community skills; the fastest-growing repo in GitHub history.
  2. Hermes (247k stars, slightly complex, autonomy headless, recovery resumable): Nous Research's self-improving agent: a learning loop turns experience into reusable skills, builds a persistent user model across sessions, and checkpoints state to disk with rollback; lean enough for a $5 VPS, driven from chat, and model-agnostic (Nous Portal, OpenRouter, OpenAI, or any endpoint).
  3. AnythingLLM (66.3k stars, complex, autonomy headless, recovery resumable): Self-hosted "AI second brain" harness: chat with your documents, run built-in agent skills (web search, code execution, browsing), and manage multi-user workspaces with a bundled vector DB, so there is no separate RAG stack to wire up.

Frameworks

General-purpose agent and LLM application frameworks (the app layer, not harnesses per se). Full category →

  1. n8n (205k stars, complex, autonomy headless, recovery durable): Fair-code workflow engine with 400+ nodes and native AI nodes; the self-hosted Zapier that actually does agents and LangChain.
  2. AutoGPT (187k stars, complex, autonomy headless, recovery resumable): The original autonomous loop: goal in, agent iterates with tools and memory; Forge is the dev framework, Benchmark the eval harness.
  3. Dify (157k stars, complex, autonomy headless, recovery retry): One-stop LLM app platform: visual workflows, RAG pipeline, 50+ tools, model management; "ship from prototype to prod" in a single UI.

Multi-agent and orchestration

Harnesses and patterns for multi-agent coordination and handoffs. Full category →

  1. MetaGPT (70.5k stars, complex, autonomy headless, recovery resumable): The "AI software company" multi-agent framework: role-played PM, architect, and engineer agents turn a one-line requirement into specs, designs, and code along an SOP assembly line. The landmark of the genre; development pace has slowed in 2026.
  2. autogen (61.1k stars, complex, autonomy bounded, recovery resumable): Conversable agents and group chats; code execution and human-in-the-loop; Microsoft origin, AG2 ecosystem. ⚠️ In maintenance mode since late 2025: no new features, and Microsoft directs new users to Agent Framework.
  3. crewAI (58.8k stars, complex, autonomy bounded, recovery resumable): Role-based agents (roles, goals, backstories) in Crews; Flows add event-driven and hierarchical control for production.

Plugins, MCPs, CLI tools

IDE plugins, concrete MCP servers, and CLI tools that give agents tools and context. Full category →

  1. MCP Servers (90.5k stars, mostly simple, not a loop-owning harness (a format, config, or library)): The official reference collection of Model Context Protocol servers (filesystem, git, fetch, memory, time, and more)—the canonical, vetted toolset agents connect to, and the pattern every other MCP server is measured against.
  2. Context7 (62.2k stars, super simple, not a loop-owning harness (a format, config, or library)): MCP server that injects up-to-date, version-specific library docs into an agent's context on demand; kills the stale-training-data hallucinations that plague codegen.
  3. chrome-devtools-mcp (52.4k stars, mostly simple, not a loop-owning harness (a format, config, or library)): Google's official Chrome DevTools MCP server: exposes console, network, and performance-trace inspection as tool calls agents can drive directly, instead of a human clicking through the DevTools panel.

Memory and state

Persistent memory layers that give agents recall across turns and sessions: knowledge graphs, vector stores, and session-capture tools that survive a restart. The state a harness needs but rarely ships with. Full category →

  1. claude-mem (94.3k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Session-memory plugin for Claude Code, Codex, OpenClaw, Gemini, Copilot, and more: captures everything an agent does during a session, AI-compresses it, and injects the relevant context into future sessions. Session-to-session memory as a drop-in.
  2. Mem0 (65.7k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Universal memory layer for AI agents: stores user/org/session memory, retrieves on demand. Apache-2.0; the de-facto memory primitive paired with most harnesses in 2026.
  3. Graphiti (Zep) (31k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Zep's open-source memory engine: real-time temporal knowledge graphs that track how facts about users and entities change over time, so agents can answer "what was true when." The layer behind Zep's hosted memory platform.

Evaluation and benchmarking harnesses

Agentic eval systems, reasoning benchmarks, and open agent benchmarks. Full category →

  1. Agent Lightning (18.4k stars, complex, autonomy headless, recovery resumable): Microsoft's training-oriented harness: optimization loops for agent behavior—when you need to improve policies over rollouts, not only score a fixed prompt.
  2. SWE-bench (5.9k stars, slightly complex, autonomy headless, recovery resumable): LMs resolve real GitHub issues; Docker harness, instance IDs; standard for code-agent evals.
  3. AgentBench (3.7k stars, complex, autonomy headless, recovery none): ICLR'24 benchmark: agents across AlfWorld, DB, knowledge graphs, OS, webshop; Docker Compose, function-calling interface.

Observability and eval-ops

Tracing, monitoring, and production evaluation for live agent runs: capture every step, tool call, and token, then score and debug in the loop. Distinct from the fixed-task benchmarks above—this is what you run against your own traffic. Full category →

  1. Langfuse (34.9k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Open-source LLM engineering platform: full-trace observability, online and offline evals, prompt management, and cost metrics for agent runs in production—the monitoring layer most harnesses lack out of the box.
  2. MLflow (28.1k stars, complex, not a loop-owning harness (a format, config, or library)): Mature ML platform now covering GenAI: MLflow Tracing captures every agent step, tool call, and token, with built-in LLM evals and prompt versioning—observability for teams already standardized on MLflow.
  3. Opik (22.2k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Comet's open-source agent observability and evaluation platform: tracing, scoring, and experiment comparison with the whole core feature set free to self-host under Apache-2.0.

Research and task-specific harnesses

Deep research, document QA, and domain-specific agent loops. Full category →

  1. DeerFlow (82.8k stars, complex, autonomy headless, recovery resumable): ByteDance's long-horizon research harness: a LangGraph-based agent loop that researches, codes, and creates using sandboxes, memory, tools, and subagents behind a message gateway—an alternative to the reference gpt-researcher stack for multi-hour tasks.
  2. gpt-researcher (29.5k stars, complex, autonomy bounded, recovery retry): Autonomous deep-research agent: web + local sources, citation-grounded reports, multi-agent and deep-research modes. The reference open-source research harness.
  3. AutoResearchClaw (14.5k stars, complex, not a loop-owning harness (a format, config, or library)): Fully autonomous, self-evolving research harness: multi-agent debate and citation verification carry a project from idea to a written paper without a human in the loop.

Libraries and SDKs

Lightweight runtimes, tool loops, and provider-agnostic harness primitives. Full category →

  1. Daytona (71.7k stars, slightly complex, not a loop-owning harness (a format, config, or library)): Elastic dev environments for AI-generated code: workspaces, Git, previews—infra harness between "the model wrote a patch" and "it ran in a real machine." ⚠️ Public repo unmaintained since June 2026: core development moved to a private codebase (final open release v0.190.0, AGPL-3.0).
  2. LiteLLM (59.2k stars, mostly simple, not a loop-owning harness (a format, config, or library)): One interface to 100+ LLMs; routing, caching, budgets. Not an agent framework—the pipe every agent framework uses.
  3. Composio (30.3k stars, complex, not a loop-owning harness (a format, config, or library)): 1,000+ toolkits with auth, tool search, and a sandboxed workbench—drop-in tool layer so agents stop reinventing OAuth + integrations. Python and TypeScript.

Part of best-of-Agent-Harnesses. Regenerated from the list data on every weekly refresh.

Agents can query this page's data directly: claude mcp add agent-harnesses -- uvx agent-harnesses-mcp (MCP server).