SciencePal 2.0
The Agent Evolves Itself · Multi-Agent Research Platform
SciencePal is a multi-agent research assistant developed by CityU HKAI-SCI. Through natural conversation, delegate literature review, data analysis, code execution, and report writing to domain-specialized agents running in isolated sandboxes — your trusted AI research partner.
Domain Expert Agents
Covering biology, materials, medicine, surveys, MIMIC clinical data, and more — queries are automatically routed to the best-fit agent.
Sandbox as Computer
Each task runs in an isolated Daytona sandbox with Shell, browser, file I/O, and code execution — transparent and auditable.
Self-Evolving Agents
Skills, Memory, Subagents, and knowledge graphs evolve together — agents learn your research habits over time.
SciencePal 2.0 Core Upgrades
Four capability modules — from tool use to self-growth
Skills · Reusable Know-How
- Skill-Creator builds more domain skills automatically
- Built-in bio, medical, and materials skill packs
- Pluggable — drop in and auto-discover on load
Memory · Persistent Recall
- BM25 keyword + HNSW vector hybrid retrieval
- 5 typed memory classes, injected at session start
- High-value memories can graduate into Skills
Subagent · Parallel Work
- Parent delegates to children — general, explore, plan roles
- One agent, many workers in parallel
- Max 10 concurrent, no nesting for controllability
Knowledge Graph · Your Library
- Two-pass extraction: AST structure + semantic understanding
- Confidence and provenance on every edge
- Wiki articles per community — build your research library
Platform Core Features
End-to-end capabilities for delegating, executing, and collaborating on research tasks
Domain Agents
Pre-built experts across a dozen research fields — the system matches each question to the best agent.
Quick select on workspace · @ mention to switch
Agent Co-Building
Create and configure custom agents with tailored prompts, tools, and collaboration settings.
Drafts · audit trail · team sharing
Isolated Sandbox
Each task runs in a secure, isolated environment with command execution, browsing, and file operations.
Live desktop view · save or revert changes
Personal Knowledge Base
Manage your literature and documents, linked to agent tasks for retrieval and citation.
Upload · semantic search · in-conversation citation
Skills Ecosystem
Reusable skill documents extend agent capabilities for literature, visualization, writing, and more.
On-demand loading · continuous evolution · external platforms
Self-Evolution
Agents learn and improve over time as memory and skills grow together.
Memory retention · skill upgrades · multi-agent coordination
Platform Integration
Ask SciencePal from Feishu, WhatsApp, and other tools — same quality replies as the web app.
Message sync · async delivery · consistent across devices
Observability & Collaboration
Track task progress, share results, and receive email notifications at key milestones.
Share links · execution replay · email alerts
Domain Agents Overview
Select an agent on the workspace or @ mention in chat — the system finds the right domain expert
General Agent
General-purpose agent for all types of scientific questions
Biology Agent
Focused agent for structural biology and bioinformatics analysis
Materials Agent
Specialized agent for materials science
Survey Agent
Professional academic survey agent for systematic literature review and comprehensive survey report generation
AMix Agent
Advanced AMix agent for protein engineering and molecular interaction design
Medical Agent
Structured health consultation and clinical triage guidance; emergency red flags; not for diagnosis
MIMIC Agent
Faithfully reproduce MIMIC-IV workflows from the literature with read-only SQL, statistics, figures, and full transparency
Plasma Agent
Specialized agent for plasma properties research
Antibody Agent
Specialized agent for antibody sequence analysis, developability assessment, antibody-antigen modeling, and candidate antibody design
More domain agents are coming soon — co-building welcome. Custom agents are also supported.
Product Experience Journey
Four steps from question to result
Pick an agent and describe your task
After sign-in, open the workspace, choose a domain agent, and describe your need — or start from an example prompt. Send to create a session and enter the execution view.
- Switch between agents in one click: General, Biology, Materials, Survey, and more
- Built-in example prompts to get started quickly
- Configure web search, email notifications, and other preferences
- One click to send — the system opens the task view automatically
Watch progress in real time — ask follow-ups or stop anytime
The conversation view shows the agent's reasoning on the left; the action panel on the right displays tool steps, generated files, and the remote desktop.
- Click a tool tag in a message to jump to the matching step in the panel
- Preview files as the agent writes them
- Follow up, stop the task, switch models, or enable deep thinking
- Generate a share link so others can view the task and results
The sandbox is the agent's dedicated computer
Each task runs in an isolated sandbox where the agent executes commands, browses the web, and reads or writes files — all visible in the action panel.
- Web search and academic literature retrieval built in
- Complex tasks can dispatch multiple sub-agents in parallel
- Save or discard sandbox changes before the task ends
- Watch the agent work in the sandbox in real time
History, search, and organization
The project sidebar lists all past sessions: start new chats, search past content, rename, share, or delete in bulk.
- Clear status indicators: running, completed, failed, or stopped
- Full-text search across projects with highlighted matches
- Quick access to knowledge base, skills, and platform integrations from the sidebar menu
- Keyboard shortcut to collapse or expand the sidebar
Extend SciencePal into your workflow
Core capabilities available beyond the web browser
Personal Knowledge Base
Access from the sidebar to upload and manage literature, building a personal library for agent retrieval and analysis.
Skills Package
Download the official skills package and follow the setup guide to connect OpenClaw, Claude Code, and other external agent environments.
Platform Integration
Link Feishu or WhatsApp to ask SciencePal from messaging apps with the same replies as on the web.
Session Sharing
Generate share links for read-only viewing of the full conversation and execution — ideal for demos, reports, and teamwork.
Ready to get started with your research?
Describe your question in natural language — SciencePal automatically routes to the right agent for search, analysis, and reporting.
SciencePal 2.0 — The Agent Evolves Itself !