EvoScientist

Overview#

EvoScientist is a multi-agent collaborative AI research automation framework covering the full scientific research lifecycle: Intake → Plan → Execute → Evaluate → Write → Verify. Its core philosophy enables AI agents to autonomously explore, generate insights, and iteratively improve, while accumulating research experience through persistent memory and self-evolution mechanisms.

Multi-Agent Collaboration#

• 6 sub-agents: planner-agent, research-agent, code-agent, debug-agent, data-analysis-agent, writing-agent, coordinated via shared LangGraph state machine
• Evolution Manager Agent (EMA): extracts reusable knowledge from historical interactions to continuously optimize research strategies
• The paper (arXiv:2603.08127) describes a 3-agent architecture (RA/EA/EMA); the code has expanded to 6 sub-agents, with exact mapping relationship to be confirmed

Persistent Memory System#

• Ideation Memory: summarizes viable research directions, records failed directions to avoid repeated exploration
• Experimentation Memory: captures effective data processing and model training strategies
• Context, preferences, and experimental findings persist across sessions (SQLite session database)

Scientific Workflow Engine#

• 6-stage pipeline using baseline-first + single-variable iteration design for scientific rigor
• Sandboxed code execution: 300s timeout, output limits, auto-recovery, "More Effort" iterative refinement mode
• In-depth literature research: web search + 7-dimensional structured reflection, enforced citation standards

Multi-Provider & Multi-Channel#

• 9 LLM providers: Anthropic, OpenAI, Google, NVIDIA, SiliconFlow, OpenRouter, Volcengine, DashScope, Ollama/Custom
• CLI/TUI as core hub, supporting 10 messaging platforms: Telegram, Slack, Feishu, WeChat, DingTalk, iMessage, Discord, Email, QQ, Signal — all sharing the same agent session

Adaptive Interaction#

• Dynamically rewrites system prompts based on conversation state
• Shows only relevant tools per turn to reduce noise

Plugins & Extensions#

• MCP Protocol: dynamically add MCP servers via EvoSci mcp add command
• EvoSkills: 10 pre-built research lifecycle skills (research-ideation, idea-tournament, experiment-pipeline, experiment-craft, paper-planning, paper-writing, paper-review, paper-rebuttal, academic-slides, evo-memory), compatible with Claude Code, Cursor and other third-party AI coding agents

Framework Dependencies#

• LangChain — agent framework foundation
• DeepAgents — batteries-included agent harness
• LangGraph — state machine orchestration and multi-agent coordination
• MCP (Model Context Protocol) — external tool integration protocol
• Runtime: Python 3.11+ (< 3.14); Docker image includes Python 3.11, Node.js 24 LTS + npx

Installation#

# Recommended: uv tool install
uv tool install EvoScientist
EvoSci onboard

# pip install
uv pip install EvoScientist

# Docker
docker run -it --rm --env-file .env -v "$(pwd)/workspace:/workspace" -v evosci-data:/home/evosci/.evoscientist ghcr.io/evoscientist/evoscientist:latest

Benchmark Performance & Academic Recognition#

• 🏆 ICAIS 2025 AI Scientist Track: 6/6 papers accepted, Best Paper & AI Reviewer's Appraisal Award
• 🥇 DeepResearch Bench II: Ranked #1
• 🥇 DeepResearch Bench: Ranked #1
• 🥇 AstaBench Code & Execution: Ranked #1
• 🥇 AstaBench Data Analysis: Ranked #1

Note: Specific benchmark scores are not publicly available; only rank positions at time of submission are stated.

Pending Information#

• Exact EvoSkills installation command not explicitly documented
• Mapping between paper's 3-agent and code's 6-agent architectures not detailed
• Memory module storage format and capacity limits not disclosed
• Web UI not currently available; listed as upcoming in roadmap