Plan Agent
Retrieves workflow patterns from Plan RAG and creates a structured analysis plan.
Architecture
A crew of agents for biological data analysis.
BioMaster decomposes a user objective into a checked workflow, then executes each step with retrieval support and bounded recovery.
Execution loop
Every step must be planned, executed, debugged, and checked before the workflow advances.
Task Agent
Uses Execute RAG to write tool-specific commands and scripts for the current step.
Debug Agent
Parses runtime errors, retrieves troubleshooting knowledge, and retries corrected execution.
Check Agent
Verifies output existence, format, and completeness before downstream steps begin.
Dual retrieval
Plan RAG and Execute RAG solve different retrieval problems.
Plan RAG
Stores high-level methodological knowledge such as workflow templates, analysis strategies, tool choices, input requirements, and expected outputs.
It is consulted during planning so the system can translate a natural language goal into a coherent multi-step workflow.
Execute RAG
Stores low-level tool knowledge such as installation notes, command-line examples, parameters, configuration details, and troubleshooting hints.
It is consulted during execution and debugging so scripts are grounded in tool-specific evidence.
Implementation
Designed for long, multi-step workflows rather than one-shot command generation.
Condensed memory
Diagnostic history is summarized so later steps retain useful context without loading full logs.
Stepwise gates
A workflow does not proceed until the current step has passed execution and output checks.
Model-agnostic design
BioMaster can run with proprietary or open-source LLM backbones through configuration.