L3 · Agent Layer
Agent Design & Deployment
Multi-Agent Collaboration System
Multi-Agent collaboration system · pipeline orchestration · task scheduling · self-developed Agent framework--turning models into executable action units in specific business scenarios
WHY AGENT LAYER
Why Agents Are Needed Above Models
Models answer questions--Agents solve problems. Models are the brain, Agents are the hands and feet
Turn Models into Action Units
Models can only generate text--Agents can call tools, schedule pipelines, chain multi-step computations, turning model capabilities into executable business processes.
Multi-Agent Collaboration
Single models have limited capabilities--Protein Agent, Gene Agent, Tumor Agent each specialize, Orchestration Agent coordinates task decomposition, working like a real bioinformatics team.
Interactive Interpretation
Agents don't just output files--they interpret results, answer follow-ups, recommend next steps. Upgrading "get numbers and judge yourself" to "AI explains meaning and gives advice".
ARCHITECTURE
Four-Layer Architecture
Inter-layer MCP protocol communication--dialog -> collaboration -> tools -> computing
Dialog Interaction Layer
L3 · Layer 1Natural language -> Agent understands intent -> schedules tools -> generates reports. Continuously understanding context in multi-turn dialog, translating researchers' vague questions into precise computing tasks.
Multi-Agent Collaboration Layer
L3 · Layer 2Protein Agent, Gene Agent, Health Agent each handle their domain--Orchestration Agent decomposes tasks, distributes to domain Agents, aggregates results. Collaborating like a real bioinformatics team.
MCP Tool Layer
L3 · Layer 3Streamable HTTP exposes 8+ tools, JWT authentication, per-dev index isolation. Agents call underlying computing capabilities through standard MCP protocol--tools are hot-pluggable, capabilities are extensible.
Computing Pipeline Layer
L3 · Layer 46 pipelines, 30+ AI model-driven--Docker containerized, reproducible delivery. Every computing task scheduled by Agents runs on validated pipelines.
CAPABILITIES
What Agents Can Do
Dialog entry points for 6 pipelines--each Agent encapsulates a complete business capability
Protein Structure Prediction
Submit sequences via dialog, Agent auto-calls multi-model prediction, cross-validates confidence
Enzyme Activity Assessment
DLKcat + FoldX + triple-model cross-validation of enzyme activity changes
Gene Editing Plan
Specify target gene, Agent auto-generates gRNA, scans off-targets, optimizes codons
mRNA Design
From antigen sequence to complete mRNA design--one conversation
Precision Oncology
Submit WGS data, Agent auto-annotates -> scores -> medicates -> neoantigens
Interactive Interpretation
Agent doesn't just output files--it interprets results, answers questions, recommends plans
MCP TOOLCHAIN
MCP Toolchain
8+ tools exposed via Streamable HTTP, JWT authentication--Agents call underlying computing through standard protocol
il_query
Asset dynamic knowledge base query
divo_search
Unified hybrid search (BM25 + BGE-M3)
protein_predict
Protein structure prediction scheduling
enzyme_activity
Enzyme activity multi-model cross-validation
gene_edit_design
gRNA design and off-target scanning
mrna_design
mRNA full-module design
variant_annotation
Variant annotation and pathogenicity scoring
coverage_scanner
Pipeline coverage scanning and plan recommendation
COMPARISON
Agent vs Traditional Approach
From "submit task and wait days" to "ask questions and get real-time scheduling"
| Agent | ||
|---|---|---|
| Interaction | Submit task -> wait days -> receive file -> interpret yourself | Ask questions -> Agent real-time scheduling -> output results+interpretation+recommendations |
| Cross-pipeline collaboration | Manual chaining, each pipeline submitted separately | Agent auto-decomposes tasks, schedules multi-pipelines, aggregates results |
| Result interpretation | Get numbers and judge yourself | Agent interprets meaning, cross-validates, recommends next steps |
| Knowledge reuse | Repeat background explanation every time | RAG + Investigate Lens auto-retrieves team knowledge |
STATUS
Development Status
MCP toolchain is live, dialog engine and domain Agents continuously iterating
OPEN SOURCE PLATFORM
Agent Runtime Governance Platform
The above Agent capabilities run on a unified governance platform--AGPL-3.0 open source, community edition free
DiVo Agent Platform - Enterprise-grade AI Agent runtime governance platform
sealionking/divo-agent-pub
AGENT SHOWCASE
Agent Runtime Live Demos
Each Agent Demo is a real interactive session—click to experience dialog-based task submission, real-time pipeline scheduling, and result interpretation
⚠ Showcase is currently under development. Demos are not yet available. Our Agent engineers are working hard — stay tuned!
Protein Structure Prediction
Dialog-based sequence submission, multi-model prediction + cross-validation
Enzyme Activity Assessment
DLKcat + FoldX triple-model cross-validation
Gene Editing Plan
gRNA generation + off-target scan + codon optimization
mRNA Design
From antigen to complete mRNA full-chain design
Precision Oncology
WGS annotation→scoring→medication→neoantigen
Interactive Interpretation
Interpret results, answer follow-ups, recommend plans
Multi-Agent Orchestrator
Auto-decompose tasks, distribute to domain Agents
Dialog Engine
Multi-turn context understanding, intent recognition, slot filling
CROSS-LAYER
Cross-Layer Positioning
L3 Agent layer is the pivot--wrapping L2 models below, delivering capabilities to L4 computing above
30+ AI model foundation--protein structure, enzyme activity, variant annotation, etc. L3 Agent layer wraps these models, exposing their capabilities as callable tools.
View Model LayerMCP + multi-Agent + RAG--wrapping model capabilities into dialog-based business units, orchestrating tasks, scheduling tools, interpreting results.
6 pipelines, Docker containerized, reproducible delivery. Every computing task scheduled by L3 Agents is actually executed and delivered by the L4 computing layer.
View Computing LayerUSER SCENE
Who These Agents Ultimately Serve
B2B2C: Each Agent capability, through dry-computing pipelines, ultimately reaches these end-user scenarios
Experience DiVo Bio Agent Platform
MCP toolchain is live, dialog engine and domain Agents under continuous development. Join us in encapsulating bioinformatics team capabilities into dialog-based interactions.