Best of Agent Harnesses
A curated, weekly-rescored ranked list of 100+ AI agent harnesses — plus an MCP server, llms.txt and JSON so agents can recommend harnesses themselves.
CI declares least privilege (contents: read) and installs only a few dependencies (pytest, markdown); data flow is reasonably transparent; source attribution is excellent with citations to Willison, Karpathy, the SWE-agent paper, etc. Deductions: the README's curl-download of agent skeletons into .claude/agents/ imports external instruction files without advising user review, thinning user confirmation; the weekly-rescore workflow pushes with GITHUB_TOKEN but its script is not in evidence; rollback relies on implicit git with no explicit guidance.
A pytest suite covers regression scenarios for the ARCHIVED block, failing deliberately via SystemExit on format drift (clear failure behavior); CI runs on PRs/pushes including an end-to-end site build. Deductions: no broader tests over the recommendation logic or rendered output; dependencies are unpinned (pip install markdown).
Clear multi-audience scenarios (human readers, coding agents, researchers) with a use-case index, capability/autonomy/recovery tiers, decision guides, and multiple MCP entry points (recommend, pick_harness). Deductions: environment fit is stated (uv, Python 3.12) but boundary conditions — offline use, rate limits, degraded behavior of pick_infrastructure when GitHub API fails — are not described.
Excellent information architecture: table of contents, tag cross-references, tiered tables, FAQ, machine-readable artifacts (llms.txt/JSON); full CC-BY-SA-4.0 license text present. Deductions: no CHANGELOG or formal versioning (only a stars-capture date); naming stability rests on a single maintainer; the maintenance claim ('merged fast') is asserted rather than institutional.
High output usability: site, JSON, MCP, and agent skeletons give multiple consumption forms; clear marginal value — solving harness selection for a given model/task — refreshed weekly; cost is minimal (reading a list or one MCP call), so cost-benefit is strong. No specific deduction found; full marks are supported by the text evidence.
Strong claim traceability: ranking methodology, stars-capture date, and external benchmark citations (SWE-bench Pro, AINews) are linked; editorial scores are explicitly labeled and open to maintainer correction. Deductions: some citations (2026-dated sources, forward-looking benchmark figures) cannot be independently corroborated in a static review; fact/inference separation is only partially achieved (scores labeled, but numeric star counts unverifiable).
- curl-installing the agent skeletons places remote Markdown instructions directly into .claude/agents/; review the file contents manually before installing.
- The weekly-rescore workflow pushes with GITHUB_TOKEN, but its script does not appear in the reviewed evidence; verify its behavior.
- Star counts, dates (including 2026), and external citations in the README are static text; check freshness before relying on them.
- Publisher identity is unverified in the curated registry; trust recommendations based on in-repo evidence, not identity.
- Dependencies are unpinned (pip install markdown/pytest); pin them yourself when reproducing builds.
What does this agent do, and when should you use it?
This repository is a curated, ranked, weekly-rescored list of AI agent harnesses — 160+ projects spanning coding agent products, frameworks, multi-agent orchestration, memory layers, sandboxed execution, and evaluation harnesses. Each entry carries stars, open-source status, a simplicity-to-capability tier, autonomy and recovery tier markers, and one concrete example link. The list is also published in machine-readable form as harnesses. and llms.txt, and it ships an MCP server (published to PyPI and the official MCP registry as io.github.RyanAlberts/agent-harnesses) exposing tools like recommend, pick_harness, and compare. The repo additionally bundles three installable agent skeletons — harness-scout, stack-auditor, and harness-radar — that run on an existing AI subscription and deliver results to Slack or Notion. Licensed CC-BY-SA-4.0, it also offers a searchable web site filterable by capability, autonomy, and recovery.
Weekly, the repository rescrapes and reranks 160+ harness projects, scoring them by harness concerns (environment, orchestration, lifecycle, guardrails) plus stars/activity, and auto-regenerates landscape charts (landscape.svg, axes-grid.svg) from the list data. It publishes three machine-readable artifacts: harnesses. (category, complexity tier, capability tags, stars, license signal, example link per project), llms.txt (the entire list in one agent-readable file), and an MCP server whose tools include recommend (one opinionated pick + alternatives + what to avoid, e.g. repos flagged for star manipulation), compare/compare_for (2–4 harnesses side by side across axes), pick_harness (ranked picks with complexity/autonomy/recovery filters), pick_infrastructure (picks at any infra level plus a live GitHub/Hacker News discovery pass), search_harnesses, get_harness, list_categories, list_comparisons, and get_comparison. The agents/ directory ships three skeletons: harness-scout (picks a harness for what you're building, with evidence and a graveyard check), stack-auditor (flags harnesses in your codebase that died), and harness-radar (weekly movement briefing: climbers, arrivals, deaths, graduations), all delivering to Slack or Notion. The list also includes comparison decision guides such as how to pick a harness, a two-week trial protocol, terminal coding agent head-to-heads, and multi-agent orchestration comparisons.
- A team choosing a coding harness for their model and task wants benchmark-backed evidence rather than defaulting to whatever someone else benchmarked — harness swaps moved pass@1 from 23% to 52% on SWE-bench Pro
- An agent developer wants their coding agent to call MCP tools (recommend, pick_harness) to auto-recommend a matching harness for users
- An engineering lead wants to compare terminal coding agents (opencode, Codex, Gemini CLI, crush, goose) or orchestration frameworks (OpenAI Agents SDK, CrewAI, AutoGen, LangGraph)
- A team needs to select an agent memory layer (Mem0, Letta, claude-mem), sandboxed execution (E2B, Daytona, Modal), or eval/observability platform (Langfuse, LangSmith, Braintrust, Phoenix)
- An engineer maintaining a multi-agent stack wants a weekly briefing of harness climbers, arrivals, deaths, and graduations (harness-radar skeleton)
- An individual wants an always-on personal agent in chat apps and can filter the personal agent runtimes category (OpenClaw, Hermes, Khoj, Agent Zero, etc.)
What are this agent's strengths and limitations?
- Scale and maintenance cadence: 164 projects, rescored weekly, with a searchable site and auto-regenerated landscape charts
- Natively agent-facing: an MCP server (on PyPI and the official MCP registry), llms.txt, and harnesses. let agents query recommendations directly rather than only humans browsing
- Evidence-based selection content: cites SWE-bench Pro data showing harness choice moves scores more than model upgrades (23% to 52% pass@1 on the same model), plus a two-week trial protocol and graveyard warnings (repos flagged for star manipulation)
- Three ready-to-use agent skeletons (harness-scout, stack-auditor, harness-radar) clone straight into .claude/agents/ and support Slack/Notion delivery
- Scores are editorial, derived from public docs rather than first-hand testing; maintainers must self-correct via issue/PR
- The README is truncated relative to the full list, and some entries — especially low-star projects — carry ❓ (no or unclear) license signals requiring verification before adoption
- Several listed projects carry restrictive or source-available licenses (⚠️ Fair-code, FSL-1.1-MIT, Elastic-2.0, Polyform-SU, etc.) and cannot be treated as plain open source
- Star counts are a snapshot (captured 2026-09-09), and the project is an ecosystem information aggregation, not itself a runnable agent product beyond the MCP server and skeletons
How do you install or deploy this agent?
The repository itself is a list/data project — clone it: git clone https://github.com/RyanAlberts/best-of-Agent-Harnesses. Machine-readable data is available directly at the raw URL https://raw.githubusercontent.com/RyanAlberts/best-of-Agent-Harnesses/main/llms.txt and harnesses. in the repo. The MCP server is published to PyPI and the official MCP registry (io.github.RyanAlberts/agent-harnesses); install uv first, then run: claude mcp add agent-harnesses -- uvx agent-harnesses-mcp. Example skeleton install: curl -fsSL https://raw.githubusercontent.com/RyanAlberts/best-of-Agent-Harnesses/main/agents/harness-scout.md -o .claude/agents/harness-scout.mdHow do you use this agent?
Three main usage modes. 1) Human readers: browse the searchable site at https://ryanalberts.github.io/best-of-Agent-Harnesses/, filter by capability, autonomy, and recovery, or use the "Pick by use case" index and comparison decision guides. 2) Agent queries: after registering the agent-harnesses MCP server in an MCP-capable client, agents can call recommend, pick_harness, compare, pick_infrastructure, and related tools for evidence-backed harness recommendations. 3) Skeleton agents: drop the harness-scout, stack-auditor, or harness-radar markdown files from agents/ into .claude/agents/, customize the instructions as needed; with Slack or Notion connected, the skeletons deliver results each data cycle. All three modes operate on the current week's rescored data.
How does this agent compare with similar options?
Typical awesome-lists are statically maintained; this repo differentiates via weekly rescoring, machine-readable publication (JSON/llms.txt), and a built-in MCP recommendation server. The list itself systematically compares real alternatives across categories: opencode vs Codex vs Gemini CLI vs crush vs goose (terminal coding agents), OpenAI Agents SDK vs CrewAI vs AutoGen vs LangGraph (multi-agent orchestration), Mem0 vs Letta vs claude-mem (memory layers), E2B vs Daytona vs Modal (sandboxing), and Langfuse vs LangSmith vs Braintrust vs Phoenix (eval/observability platforms).