32 projects. 11 categories. Opinionated curation. Contains no original code — all repos are forks/clones for analysis. See Local mirror to reproduce the bench.
Project
Stack
Differential
anthropics/claude-code TS + Ink + Bun
Snapshot (~512K LOC) of Anthropic's official CLI; canonical reference for terminal-UI coding agents.
openai/codex TS + Rust
OpenAI's official CLI coding agent with IDE integrations and a focus on local automation.
Aider-AI/aider Python
Most-used pair-programming CLI in open source; git-native with auto-commits per iteration.
All-Hands-AI/OpenHands Python
Most-starred autonomous agent on GitHub (formerly OpenDevin); standard SWE-bench benchmark.
Project
Stack
Differential
gsd-build/gsd-2 TS (Pi SDK)
Auto-milestones with no human in the loop; RTK compresses shell output across long runs.
garrytan/gstack TS + Playwright
Personal software factory by YC's president; 24 specialized agents in a multi-role workflow.
obra/superpowers TS (Claude Code plugin)
Pure red/green TDD with parallel subagents; reports a 94% PR rejection rate.
🤝 Multi-agent orchestration
Project
Stack
Differential
bmad-code-org/BMAD-METHOD TS
12+ personas (PM, architect, UX…) with "Party Mode" running multiple in one session; agile applied to agents.
crewAI-inc/crewAI Python
Framework for collaborative role-playing agents with shared goals.
microsoft/autogen Python
Microsoft's multi-agent conversational framework; research baseline for message-based coordination.
paperclipai/paperclip TS + React + Postgres
Manages agents like employees — org charts, budgets and goals.
grandamenium/claude-remote-manager Bash + TS
Claude Code 24/7 controlled via Telegram; persistent cron survives restart.
Project
Stack
Differential
mem0ai/mem0 Python
Leading memory layer in 2025; native vector store and a simple API for LLM apps.
milla-jovovich/mempalace Python + Chroma
96.6% R@5 with verbatim storage (no paraphrasing); local-first, no API keys.
garrytan/gbrain TS + PGLite + pgvector
95% recall@5; entity self-wiring without an LLM; 30-minute setup.
MemoriLabs/Memori Python
LLM-agnostic, agent-native memory infrastructure; 81.95% on LoCoMo, SQL-backed.
Project
Stack
Differential
openclaw/openclaw TS + SwiftUI + Kotlin
20+ channels (WhatsApp, Telegram, iMessage…) running on-device.
NousResearch/hermes-agent Python asyncio
Auto-creates skills; runs on a $5 VPS or serverless with hibernation; multi-model.
📐 Spec-driven & methodology
Project
Stack
Differential
github/spec-kit Python + TS
Executable specs that generate implementation; GitHub's official methodology.
gsd-build/get-shit-done TS
Solves "context rot" via spec discipline and meta-prompting; in use at Amazon, Google and Shopify.
Systems that iterate, mutate and optimize against a metric — descendants of karpathy/autoresearch and the "AI scientist" lineage. See also alvinreal/awesome-autoresearch
Project
Stack
Differential
ShengranHu/ADAS Python
Automated Design of Agentic Systems (ICLR 2025); meta-agents that invent novel agent architectures by programming them in code.
SakanaAI/AI-Scientist-v2 Python
Workshop-level autonomous scientific discovery via agentic tree search; removes v1's template dependency and generalizes across domains.
gepa-ai/gepa Python
GEPA (Genetic-Pareto) — ICLR 2026 Oral; reflective prompt evolution that outperforms RL (GRPO); optimizes any textual parameter against any metric via natural-language reflection.
⚙️ Workflow & durable execution
Project
Stack
Differential
vercel/workflow TS + Next.js + PG
Deterministic replay via event log; split VM + step runtime.
github/gh-aw Go + Markdown
Agentic workflows written in natural language, executed sandboxed inside GitHub Actions.
🔌 Protocol & infrastructure 📊 Evaluation & observability
Project
Stack
Differential
langfuse/langfuse TS + Next.js
Most popular open-source observability for LLM apps; tracing, eval and prompt management.
Project
Stack
Differential
dataelement/Clawith TS
"OpenClaw for teams": digital employees with soul.md + memory.md, org chart and multi-tenant delegation.
microsoft/agent-governance-toolkit Multi-lang
Sub-ms policy enforcement covering 10/10 of the OWASP Agentic Top 10; runtime security.
langgenius/dify Python + TS
Most popular low-code platform (129k+ stars); ready for teams and production.
Project
Stack
Differential
alchaincyf/huashu-design Skill (multi-agent)
Agent-agnostic skill (Claude Code, Cursor, Codex, OpenClaw, Hermes) that delivers ready-to-ship design — animations, clickable prototypes, slide decks, infographics — from a single prompt.
Cross-cutting axis, grouping harnesses by recurring technique/architecture:
Pattern
Projects
Note
First-class memory gbrain, mempalace, gsd-2, hermes-agent, mem0, memori-labs
Memory as a separate, measurable component — not a bolt-on.
Spec-driven spec-kit, get-shit-done, superpowers
Pragmatic alternative to "vibe coding"; specs drive execution.
Multi-persona / party-mode BMAD-METHOD, crewAI, autogen, paperclip, gstack
Coordination across multiple roles inside a single harness.
Parallel subagents superpowers, ai-website-cloner-template, gstack
Task fan-out to specialized agents.
Durable / replay vercel/workflow, gh-aw
Event-sourced; survives crashes; reproducible.
Local-first mempalace, openclaw, gbrain, hermes-agent
No cloud dependency; on-device or self-hosted.
Governance / policy agent-governance-toolkit, Clawith
Runtime security and agent org charts.
24/7 persistent claude-remote-manager, hermes-agent, gsd-2
Cron, hibernation, automatic resume.
Self-improvement / autoresearch ADAS, AI-Scientist-v2, gepa, superpowers
A loop that measures, mutates and optimizes — code, prompts or architecture.
To clone/update every project in parallel:
The script reads repos.tsv, clones whatever is missing and runs pull --ff-only on the rest.
Legend: [+] cloned · [↑] updated · [=] up-to-date · [x] error
PRs welcome. To add a harness:
Add an entry to repos.tsv in the form name<TAB>url<TAB>branch.
Add the row to the matching section's table in the README: [owner/repo](url) | Stack | One-sentence description with the technical differential.
If it represents a new recurring pattern, add it under 🗒️ Patterns observed .
Inclusion criterion : the project must be a harness — something that orchestrates, executes, or gives memory/tools to an LLM. Pure model libraries (no agent loop) and generic infrastructure (raw vector DBs, etc.) are out of scope.
This list is released under CC0-1.0 .
