A curated guide for LLM-agent-driven scientific research automation
🌐 View Interactive Multilingual README →
🇨🇳 中文 · 🇺🇸 English · 🇰🇷 한국어 · 🇯🇵 日本語 · 🇩🇪 Deutsch · 🇫🇷 Français · 🇪🇸 Español · 🇮🇹 Italiano · 🇵🇹 Português · 🇸🇦 العربية · 🇹🇭 ไทย · 🇻🇳 Tiếng Việt · 🇷🇺 Русский
Automate the research loop with LLM agents: literature review → idea generation → experiment execution → paper writing → peer review.
This repo is a landing page for the field: use it to choose the right track, then move into the topic pages for detail.
Start here: Getting Started · Tools & Platforms · Claw Park · Vibe Coding
|
Core Question How far can AI move from research assistant to research operator? Focus: literature, ideation, experiment, writing, and evaluation. |
What Changed In 2026 Research copilots got stronger, learning layers became real, autonomous research systems got more credible, and Vibe Coding became the execution layer. |
How To Use This Repo Treat the README as a map. Treat the topic pages as the actual guide. |
Four trends are now shaping the field:
- Research copilots are getting stronger: Deep Research products, NotebookLM-style source-grounded reading, and scientific workspaces such as Prism are making literature synthesis and report writing much faster.
- Learning and self-evolving agents are becoming a real layer: Agent Lightning, Agent0, AgentEvolver, EvoAgentX, and memory substrates such as Acontext suggest that training, optimization, and persistent reusable experience are becoming first-class agent infrastructure.
- Autonomous research systems are maturing: AI Scientist-v2, Agent Laboratory, and EvoScientist push the field from "paper summary bots" toward iterative ideation, execution, and evaluation.
- Vibe Coding is becoming the execution layer: terminal agents, coding agents, and background agents now matter because research automation increasingly depends on reliable code generation, experiment loops, and repository operations.
This guide keeps Vibe Research as the core topic, then adds separate sections for Vibe Coding and Vibe Anything so the repo can expand without losing scope.
Several current signals make the field feel less like a loose collection of demos and more like an emerging stack:
- Learning and RL are becoming a first-class layer: Agent Lightning turns arbitrary agents into trainable systems with RL, automatic prompt optimization, and SFT; Agent0 and AgentEvolver push zero-data and self-generated evolution; EvoAgentX and EvoScientist push evolving workflows and scientist loops.
- Skill and memory are becoming learning substrates: Acontext and anthropics/skills suggest that reusable skills, persistent context, and accumulated experience are becoming part of the agent learning stack rather than only convenience features.
- OpenClaw is becoming a platform layer: it now reads more like a self-hosted gateway plus control UI plus skill registry plus compatible plugin-bundle layer than a single assistant app. See OpenClaw, ClawHub, and Compatible Bundles.
- FutureHouse and Edison show platformization: FutureHouse Platform, Robin, BixBench, Edison, and Kosmos show how the field is moving from repos and papers to persistent public or commercial platform surfaces.
- Execution and connectors remain the glue: MCP registries, routing layers, plugin bundles, and research connectors still determine whether these agents can actually plug into code, literature, chat surfaces, and scientific databases.
This is the layer the guide used to underweight: not just tool-using agents, but agents that can be trained, optimized, or improved over time.
| Layer | Representative resources | Why it matters |
|---|---|---|
| Agent training / optimization | Agent Lightning | Brings RL, automatic prompt optimization, and SFT to arbitrary agent systems with near-zero code changes |
| Zero-data self-evolution | Agent0 · AgentEvolver | Shows how agents can generate tasks, feedback, and training signals without human-curated data pipelines |
| Evolving workflows | EvoAgentX · EvoScientist · MetaClaw | Shifts the focus from optimizing one prompt to evolving whole workflows, skill graphs, or scientist loops |
| Skill & memory substrate | Acontext · anthropics/skills | Makes skills, context, and reusable experience part of the learning layer |
| Landscape map | Awesome-Self-Evolving-Agents | Best current GitHub-native overview of the optimization, evolution, and lifelong-agent literature |
The execution layer is still where the fastest open-source productization is happening:
- Skills marketplaces are real now: anthropics/skills, wshobson/agents, and ClawHub make reusable agent capabilities browsable and installable.
- Workflow structure is becoming explicit: SuperClaude Framework and claude-task-master show how people are layering conventions, commands, and task systems on top of coding agents.
- Routing and agent-ops are becoming infrastructure: claude-code-router, Claude Squad, and Repomix highlight provider routing, multi-agent management, and codebase packaging as real operational layers.
|
🟢 New to Vibe Research Start: Getting Started Then: Tools & Platforms |
🔵 Developer / Builder Start: Tools & Platforms Then: Vibe Coding · Systems · Experiment |
|
🔴 Researcher Start: Surveys Then: Ideation · Benchmarks |
🟣 Creator / Operator Start: Vibe Anything Then: Vibe Coding · Tools & Platforms |
Only have 5 minutes? Install InnoClaw and try it out.
| Layer | Representative projects | Why it matters |
|---|---|---|
| Research copilots | OpenAI Deep Research · Gemini Deep Research · NotebookLM · Prism | Fast literature synthesis, source-grounded reading, and scientific writing assistance |
| Research systems | InnoClaw · ResearchClaw · FARS · AI Scientist · Agent Laboratory · EvoScientist | End-to-end research assistance, automation, and experiment execution |
| AI scientist platforms | FutureHouse Platform · Robin · Edison Scientific · Kosmos | Shows the field moving from paper demos to persistent web/API platforms and validated scientific workflows |
| Learning / self-evolving layer | Agent Lightning · Agent0 · AgentEvolver · EvoAgentX · Acontext | Turns agent training, self-generated data, evolving workflows, and persistent skill/context memory into a real stack layer |
| Claw ecosystem | OpenClaw · ScienceClaw · MetaClaw · AutoResearchClaw | Foundation, specialization, online learning, autonomous pipelines, and the growing gateway / skill-distribution layer |
| Execution layer | Claude Code · Codex · Cursor Background Agents · GitHub Copilot Coding Agent | The coding and repo workflow layer that increasingly powers research execution |
| Claude Code ecosystem | anthropics/skills · wshobson/agents · SuperClaude Framework · claude-code-router | Shows how the Claude Code layer is expanding into skills, marketplaces, meta-frameworks, and routing infrastructure |
| Adjacent prompt-native tools | v0 · Lovable · Replit Agent | Useful for prototyping, but not the core of Vibe Research |
| → Tools & Platforms | → Claw Park | → Vibe Coding | → Vibe Anything |
A new layer is forming between "agent" and "workflow": plugin surfaces, MCP registries, skill catalogs, and chat bridges that make research agents easier to extend, discover, and operate.
| Layer | Representative resources | Why it matters |
|---|---|---|
| Bridge & control surfaces | cc-connect | Runs Claude Code, Cursor, Gemini CLI, Codex, and similar agents from chat surfaces such as Feishu/Lark, Slack, Telegram, and WeCom |
| Plugin / customization layer | ClawHub · OpenClaw Plugin Bundles · awesome-claude-code-plugins | Shows how agent ecosystems are moving toward skill registries, plugin marketplaces, bundle compatibility, and installable capability packs |
| Learning / memory substrate | Acontext · anthropics/skills | Shows how context, memory, and reusable skills are turning into persistent substrates for agent improvement |
| Claude Code workflow layer | wshobson/agents · SuperClaude Framework · claude-task-master | Shows how commands, agent teams, skills, and task systems are turning Claude Code into a fuller development environment |
| Routing / agent-ops layer | claude-code-router · Claude Squad · Repomix | Highlights provider routing, multi-agent session management, and codebase packaging as new operational layers around coding agents |
| Registry / discovery layer | Official MCP Registry · awesome-mcp-servers · awesome-openclaw-skills | Makes it easier to find, compare, and install the rapidly growing tool and skill ecosystem |
| Research connectors | OpenAlex Research MCP · Academia MCP · PapersWithCode MCP | Connects agents directly to literature graphs, code artifacts, datasets, and benchmark metadata |
More detailed map: → Tools & Platforms
| Topic | Description | Link |
|---|---|---|
| 🚀 Getting Started | 5-min demo → 30-min agent deployment → full automation | → Getting Started |
| 🧰 Tools & Platforms | Core platforms, literature tools, writing aids, experiment tools | → Tools & Platforms |
| 🦞 Claw Park | Ecosystem map for what each Claw project is building and where it fits | → Claw Park |
| 💻 Vibe Coding | Terminal agents, coding agents, background agents, and repo guardrails | → Vibe Coding |
| 🎨 Vibe Anything | Adjacent prompt-native workflows for apps, design, writing, slides, and ops | → Vibe Anything |
| Topic | Core Question | Papers | Link |
|---|---|---|---|
| 📄 Surveys | Landscape & evolution of the field | 5 | → Surveys |
| ⚙️ Systems | How to design end-to-end research systems | 6 | → Systems |
| 💡 Ideation | Can LLMs generate novel ideas | 6 | → Ideation |
| 📚 Synthesis | How to synthesize literature at scale | 5 | → Synthesis |
| 🧪 Experiment | How agents automate experiments | 4 | → Experiment |
| ✍️ Writing & Review | LLM-assisted writing & peer review | 4 | → Writing & Review |
| 📊 Benchmarks | How to evaluate research agents | 5 | → Benchmarks |
|
Read The Field Surveys · Systems · Benchmarks |
Build The Stack Tools & Platforms · Claw Park · Vibe Coding |
Prototype Beyond Research Vibe Anything |
Introductions: AI for Science (Nature) · LLM Agents (Lilian Weng) · Agentic Patterns (Andrew Ng)
Awesome Lists: LLM Agent Survey · AI Agents · Scientific Idea Generation
Search & Reading: Semantic Scholar · Elicit · Consensus · Connected Papers
AI Scientist Platforms: FutureHouse Platform · Robin · Edison Scientific · Kosmos
Learning / Self-Evolving Agents: Agent Lightning · Agent0 · AgentEvolver · EvoAgentX · Acontext · Awesome-Self-Evolving-Agents
Execution: Claude Code · Codex · Cursor Background Agents · GitHub Copilot Coding Agent · Gemini CLI
Claude Code Ecosystem: anthropics/skills · wshobson/agents · SuperClaude Framework · claude-code-router · Claude Squad · claude-task-master · Repomix
Prototyping: v0 · Lovable · Replit Agent · Figma AI · Canva AI
Conferences: NeurIPS · ICML · ICLR · ACL · AAAI · EMNLP
Submit resources via Resource Suggestion · Contribute via PR · Follow the curation guidelines
Citation
@misc{viberesearch2026,
title = {Vibe Research Guide},
author = {Aaron Wang and Contributors},
year = {2026},
url = {https://github.com/SpectrAI-Initiative/Vibe-Research-Guide},
}
Changelog
- 2026-W16: Added a dedicated learning / RL / self-evolving layer to the guide, including Agent Lightning, Agent0, AgentEvolver, EvoAgentX, Acontext, and Awesome-Self-Evolving-Agents
- 2026-W14: Added 2026 Q1 signals for OpenClaw platformization, FutureHouse / Robin / BixBench, and Edison Scientific / Kosmos; refreshed ecosystem framing across the guide
- 2026-W14: Added recent Claude Code ecosystem signals, including anthropics/skills, wshobson/agents, SuperClaude, claude-code-router, Claude Squad, claude-task-master, and Repomix
- 2026-W13: Added a new plugin / bridge / registry layer to the guide, including cc-connect, OpenAlex Research MCP, Academia MCP, PapersWithCode MCP, and more Claw ecosystem positioning
- 2026-W13: Added core tools & platforms (InnoClaw, ResearchClaw, FARS, Orchestra, OpenClaw, EvoScientist); added Deep Research tools, OpenAI Prism, MCP Servers; switched all content to English; expanded to 35+ papers across 9 topic files
- 2026-W12: Redesigned README into a stronger landing page with cleaner hierarchy, card-style path selection, and a more visual ecosystem map
- 2026-W12: Hub-and-spoke architecture reorganization
- 2026-W12: Initial public release
Full history: CHANGELOG.md
MIT License ·