EvoScientist is an agent system for research work. A team of six sub-agents (plan, research, code, debug, analyze, write) moves through a scientific workflow of intake, plan, execute, evaluate, write, and verify. The project describes its model as human-on-the-loop: the agent works autonomously as a research partner and co-evolves with the researcher rather than pausing for approval at each step. It is built on Deep Agents and LangGraph, and a technical report is published on arXiv.
Memory is distilled automatically each turn and self-links into a knowledge graph (complements, contradicts, supersedes) that persists across sessions; before a task, agents run a preflight search over stored observations. AutoSkills distills recurring patterns from that memory into reusable skills on a schedule and proposes them for review through /autoskills. Per-turn tool selection keeps only relevant tools visible, and the system prompt is rewritten based on conversation state.
The CLI is the hub: a TUI by default, a classic CLI, a browser workspace via --ui webui, a headless serve mode for messaging channels, deploy for a standalone LangGraph server used by external UIs or SDK clients, and --output-format stream-json for a line-delimited JSON event stream. Telegram, Slack, Discord, Feishu, WeChat, QQ, and iMessage (macOS only) share one agent session, and a Docker image bundles the channel dependencies and sandboxes shell access to mounted volumes. Python 3.11 or later (below 3.14) is required.
Providers include Anthropic, OpenAI, Google, MiniMax, and NVIDIA NIM, with OpenRouter, DeepSeek, Moonshot, Zhipu, DashScope, Volcengine, Novita, Requesty, and Atlas Cloud added across releases; OAuth sign-in is available for CLI coding-agent subscribers. Shell commands require approval by default, with auto-approve, a command-prefix allow list, and a dangerous mode that lifts the workspace sandbox. The backend API is unauthenticated and binds to loopback by default.
Reported benchmark placements at submission time: #1 on DeepResearch Bench, DeepResearch Bench II, AstaBench Data Analysis, and AstaBench Code & Execution, and #2 overall on ResearchClawBench (Agent Mode).
Features
- Multi-agent team: six sub-agents for planning, research, coding, debugging, analysis, and writing
- Self-evolving memory: auto-distilled observations linked into a cross-session knowledge graph
- AutoSkills: reusable skills drafted from recurring memory patterns and proposed via /autoskills
- Scheduled tasks: cron-style recurring runs via /schedule or natural language, with shell-access gating
- Agent teams: installed expert skills invited into a session with /expert for consults, parallel panels, or background jobs
- Multi-provider: Anthropic, OpenAI, Google, MiniMax, NVIDIA, and OpenRouter-hosted models switched from one config or in-session /model
- Multi-channel: Telegram, Slack, Discord, Feishu, WeChat, and QQ sharing the CLI session
- MCP and skills: MCP servers added with EvoSci mcp add; skills installed from a path or GitHub
- Code generation modes: a More Effort mode that refines generated code iteratively
- Multimodal input: images and PDF or document flattening, with text-only model fallback
- Human-in-the-loop: unified approval across the main agent and sync and async sub-agents, plus an ask_user tool for clarifications
- Script access: EvoScientist_agent importable as a LangGraph graph from Python
