EvoScientist is a multi-agent research harness built on Deep Agents and LangChain for what it calls vibe research: an AI scientist that plans, researches, codes, debugs, analyses, and writes, and that improves across runs. It takes a human-on-the-loop stance, where the person reviews and steers rather than approving each step, and the agent accumulates methods, skills, and memory over time. Reported placements at submission time include first on DeepResearch Bench and DeepResearch Bench II, first on the AstaBench Code & Execution and Data Analysis tracks, and second overall on ResearchClawBench.
Six sub-agents (plan, research, code, debug, analyze, write) cooperate inside one session. EvoMemory distills each turn into observations that self-link into a knowledge graph persisting across sessions, and AutoSkills proposes reusable skills from recurring memory clusters for the user to review. Per-turn tool selection shows only relevant tools, and the system prompt is rewritten as the conversation state changes. Human-in-the-loop approval covers the main agent and sync and async sub-agents, and the agent can ask for clarification itself.
The CLI (a TUI by default, or a classic mode) is the hub. The same agent session is reachable from a desktop WebUI, a headless serve mode, messaging channels (Telegram, Slack, Feishu, WeChat, QQ), a standalone LangGraph server for external UIs and SDK clients, and a line-delimited JSON event stream for programmatic clients. Providers include Anthropic, OpenAI, Google, MiniMax, NVIDIA, and OpenRouter, with OAuth sign-in for CLI coding-agent subscribers including Codex, and recent releases add models from DeepSeek, Zhipu, Qwen, Moonshot, Xiaomi, and xAI. MCP servers and skills from GitHub are added on the fly, scheduled tasks run on a cron expression with an optional reviewer rubric, and the companion EvoSkills collection packages the research lifecycle as installable skills that also work in other CLI coding agents.
Features
- Multi-agent team: six sub-agents (plan, research, code, debug, analyze, write) in one session, with async sub-agents following the caller's model
- EvoMemory: observations distilled every turn into a self-linking knowledge graph that persists across sessions, with a per-task retrieval preflight
- AutoSkills: recurring patterns in memory drafted into reusable skills on a schedule and proposed for review
- Scientific workflow: intake, plan, execute, evaluate, write, and verify, with 200+ built-in skills including research-lifecycle skills
- Providers: Anthropic, OpenAI, Google, MiniMax, NVIDIA, and OpenRouter in one config, plus OAuth sign-in for CLI coding-agent subscribers
- Surfaces: TUI, classic CLI, desktop WebUI, headless serve, a standalone LangGraph server, and a JSONL event stream for programmatic clients
- Channels: Telegram, Slack, Feishu, WeChat, and QQ share the agent session with the CLI, several at once
- Scheduled tasks: cron-style schedules with an optional acceptance rubric graded by a separate reviewer model, with one re-run on failure
- Human-on-the-loop controls: action approval across main and sub-agents, agent-initiated clarification, and a dangerous mode for real-filesystem access with safety checks
- Adaptive context: per-turn tool selection, dynamic system-prompt rewriting, context-retry middleware, and a More Effort code-generation mode
- MCP and skills: MCP servers added by command, skills installed from a path or GitHub, and expert skills invited into a session as agent teams
- Script inference: the agent is importable as a Python object and streamed with LangChain messages
