codex |
openai/codex |
OpenAI's local coding agent CLI |
pi |
earendil-works/pi |
Pi agent harness: coding agent, runtime, multi-provider LLM API |
SoL-Pi |
NVlabs/SoL-Pi |
Standalone extension for Pi packaging four efficiency mechanisms from auto-research: action fusion, observation packing, evidence-preserving reduction, and online context compaction; opt-in, no Pi patches (MIT) |
hermes-agent |
nousresearch/hermes-agent |
Self-improving agent by Nous Research with a learning loop |
jcode |
1jehuang/jcode |
RAM-efficient coding agent harness |
prime-agent |
PrimeIntellect-ai/prime-agent |
Self-improving RLM coding and research agent |
deepseek-harness |
deepseek-ai/deepseek-harness |
Plugin-based agent harness ("everything is a plugin") |
meta-harness |
stanford-iris-lab/meta-harness |
Stanford IRIS Lab framework for automated search over task-specific model harnesses (paper: arXiv 2603.28052) |
jit-agent |
bingreeky/JIT |
JIT-Agent — meta-agent that writes a task-specific harness on the fly ("Model-as-a-Harness"); ships 11 seed harnesses and a JIT-27B checkpoint (arXiv 2608.25593) |
headlong |
laude-institute/headlong |
Laude Institute's <10K-line Bash agent microharness with persistent agency — the agent keeps thinking between external interactions and decides when to respond |
maka |
apache/maka |
Apache Software Foundation (incubating) local-first agent workspace — desktop + CLI + eval, all sharing one Runtime Host that records model messages and tool calls as recoverable execution facts |
loop-engineering |
cobusgreyling/loop-engineering |
Patterns and tooling for designing agent loops |
loopx |
huangruiteng/loopx |
Open, provider-neutral, local-first control plane for long-horizon agents (sits on top of Codex, Claude Code, Cursor, dsh) |
herdr |
herdrdev/herdr |
Herdr — Rust terminal multiplexer built as a runtime for coding agents; hosts Claude Code / Codex / Cursor / OpenCode in persistent panes with state detection (working / blocked / idle) and session resume across disconnects; Apache-2.0, Homebrew formula |
ai-agents-the-definitive-guide |
Nicolepcx/ai-agents-the-definitive-guide |
Companion notebooks for the O'Reilly book AI Agents: The Definitive Guide; 12 chapters of Colab-ready Jupyter notebooks spanning agent foundations, architectures and patterns, planning, model choice, production contracts and tool governance, deployment, evaluation, memory, cost, and threat modeling; no license declared |
MathModelAgent |
jihe520/MathModelAgent |
Agent for mathematical modeling competitions: a staged skill pipeline (analysis/modeling, coding and visualization, drawio diagrams, paper writing, verification) drives a problem from statement to a submission-ready paper; Python backend, web frontend, Docker deploy, prebuilt macOS/Windows desktop app; free for personal use, commercial use by arrangement |
pub-dsh-privacy-router |
LYiHub/pub-dsh-privacy-router |
Host-side plugin for DeepSeek Harness (dsh) that privacy-gates local vs cloud routing — deterministic rules plus a local classifier label each turn public / sensitive / unknown, keeping sensitive and unknown turns on the local model and sending only rebuilt, whitelisted plain-text public turns to the cloud; pairs with LYiHub/pub-local-ai-discovery-server for local provider discovery; MIT |
Recuris |
Gen-Verse/Recuris |
Recursive self-improvement framework (NUS/Stanford/Oxford/Princeton, arXiv 2608.24876) that improves a long-horizon agent by evolving its memory instead of its weights or prompt — a frozen agent paired with a Skill Memory, a meta-agent that patches only the failing component from structured traces, and a deterministic validation gate over paired held-out evidence; training-free and model-agnostic, SOTA on τ²-Bench Retail and Airline; Apache-2.0 |
ai-agent-book |
bojieli/ai-agent-book |
《深入理解 AI Agent：设计原理与工程实践》(AI Agents in Depth) — open-source home of the book: full text of 10 chapters built around "Agent = LLM + context + tools", 109 companion experiments under chapter1–chapter10, community translations into 15 languages, PDF/EPUB builds and an Astro web reader; Apache-2.0 |
open-code-review |
alibaba/open-code-review |
OpenCodeReview (ocr) — Alibaba's battle-tested AI code review CLI: deterministic engineering (file selection, smart bundling into isolated sub-agent contexts, template-based rule matching, comment positioning/reflection modules) wrapped around an LLM agent with a review-tuned prompt and toolset, producing line-level comments; ocr review for diffs (workspace, branch range, single commit, resumable sessions) and ocr scan for whole-file audits, plus delegation mode where a host coding agent runs the review with no OCR API key; built-in multi-language ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI/Anthropic-compatible endpoints, MCP server, CI integrations and plugins for Claude Code / Codex / Cursor / Kimi Code / OpenCode; Go, Apache-2.0 |
Qwen-Live-Harness |
QwenLM/Qwen-Live-Harness |
Qwen-Live-Harness — open-source harness built around the Qwen Omni Realtime API (Qwen3.8 Omni Flash Realtime): a local daemon plus an Electron "Host" desktop UI add realtime audio/video interaction with screen or camera context, background task delegation to Qwen Code / Qoder CLI / Codex / Claude Code / Gemini CLI (with approval handling and progress tracking), proactive audio/visual/time-condition monitors, and local long-term memory libraries with cloud consolidation; macOS-first (Windows/Linux in progress), Apache-2.0 |
rrsi |
google-research/rrsi |
RRSI (Regularized Recursive Self-Improvement of Agent Harnesses, Google Research, arXiv 2609.24972) — evolves a frozen LLM agent's harness (prompts, control flow, tools, memory, context management, sub-agents) against a fixed evolve set while regularizing the search instead of the harness: annealed edit budget, history-conditioned proposal, stall-driven redirection to untried components, leakage-screening critic, noise-adjusted gain floor, token-cost rule and pruning of non-contributing components; one loop instantiated on three domains (terminal agent / Terminal-Bench 2.1, document work / Harvey LAB, engineering design / EngDesign), candidates drafted and evaluated in isolated git worktrees with an auditable edit history; Python, Apache-2.0 |
harness-zero |
metaevo-ai/harness-zero |
Harness-Zero (Peking University / Google / HKUST, arXiv 2609.24974) — harness distillation via agent-as-harness: a student-side harness h* (tools, middleware, skills, memory in the DeepAgents abstraction) is evolved on training tasks, then adapted into a private reference harness K (action recipes, review middleware, review guidance, failure patterns) for a harnessing agent that reviews every student response during rollouts and PASSes it or REPLACEs it with a complete response valid under the fixed minimal target harness h (one bash tool, fixed system prompt) before execution; only accepted responses enter the student-visible trajectory and SFT masks reviewer-perspective reasoning, so deployment keeps the gains under h alone with no h*, K or harnessing agent; reported +21.0 pts macro-average on Qwen3.5-9B (23.3% → 44.3%) across SpreadsheetBench / AppWorld / USPTO, distilled-under-h (44.3%) beating base-with-h* attached (41.7%), agent-as-harness 81.1% vs 78.1% code-as-harness on frontier models, and 82.3% recovery of 28 harness-exclusive behavior patterns; ships evolved harnesses in harness_bank/, Python 3.12–3.13, Docker rollouts via Harbor, Tinker SFT recipe, Apache-2.0 |
Raven |
EverMind-AI/Raven |
Raven (EverMind, arXiv 2609.33439, The Harness of Harnesses for Composable Agentic Intelligence) — a Host Agent that orchestrates built-in and third-party agents (Claude Code, Codex, OpenCode, Hermes Agent, OpenClaw, MiroThinker, GitHub Copilot, Qwen Code, CodeBuddy, Qoder, Grok Build, Kimi Code, Pi) into DAG-shaped multi-agent workflows, with harness self-evolution as the point: the separate Raven Evolver consumes Raven as a library, diagnoses failures, tests candidate harness changes against benchmarks, and retains only improvements that beat baseline in reproducible evaluations; core systems are agent orchestration, Evolver, EverOS memory (cross-session context, experience and skill recall), SkillForge (local skill libraries + EverOS + SkillHub's 114,190-skill catalog) and proactivity (event monitoring + scheduled execution); four built-in agents share the evolvable harness — Raven-Research (autonomous deep research), Raven-Code (SWE-bench Pro/Verified, SWE-Refactor, DataAgentBench), Raven-Design (slide decks, brand assets, charts, diagrams, web UI) and Raven-Oncall (unattended experiment/optimization/monitoring over hours or overnight); ships a TUI and WebUI (raven web), one-line install (Linux/macOS/WSL2 + native Windows) and Docker Compose, Python 3.12+, Apache-2.0, pre-alpha |
