50% Faster! SciOne AI Rebuilds Lab Workflows with Multi-Agent Architecture

The Amazon Bedrock multi-agent collaboration framework supports a comprehensive multi-agent system with a supervisor agent and sub-agents. SciOne AI only needs to define the functional scope and prompts for each sub-agent and the types of tasks that are assigned to the supervisor agent. The system then automatically performs task distribution, execution, and output generation upon receiving a user request, significantly streamlining the deployment and debugging processes.

Before using the multi-agent architecture, SciOne AI's single-agent architecture presented two key problems. First, with the single agent, single action group architecture, the growing number of features made intent recognition prone to ambiguity. For example, when the agent received a user request querying equipment information, the agent struggled to determine whether to retrieve data from the RAG knowledge base or access equipment records from the internal system. Such challenges became more pronounced as task complexity increased, prompting SciOne AI to adopt a more scalable architecture with clearly defined module boundaries.

Second, when handling multiple parallel tasks, the single-agent architecture made independent task management and efficient scheduling difficult. SciOne AI's solution suite includes more than ten functional modules — such as lab scheduling, equipment management, and sample preparation — each with significant functional differences and complex logic for invocation order and state management. Consolidating all operational logic into a single agent led to tight coupling between modules, limiting flexibility for feature expansion and iteration.

By adopting Amazon Bedrock multi-agent collaboration, SciOne AI has addressed the intent ambiguity and module coupling issues, laying the foundation for future product improvement and feature expansion. In the new architecture, each sub-agent has clearly defined functional boundaries, enabling independent debugging and optimization, while the supervisor agent handles tasks such as intent recognition, routing, and agent invocation and scheduling. This approach boosts maintainability and responsiveness of the overall system.

"With the Amazon Bedrock multi-agent collaboration framework, we break down functionality into dedicated sub-agents and use a supervisor agent to schedule sub-agents, implementing automated intent recognition and routing. We only need to define each agent's work scope and prompts, and the system automatically routes requests. The Amazon Bedrock multi-agent collaboration framework delivers a clearer, more scalable, and maintainable architecture, and also streamlines deployment."
— Co-founder of SciOne AI

The intelligent system SciOne AI built with the multi-agent architecture uses a supervisor agent to schedule multiple task-specific sub-agents, implementing cross-module collaboration. The design improves intent recognition accuracy and interaction efficiency. In customized scenarios, the solution allows deep integration with customer-side tools and dedicated agents, aligning product capabilities with unique business needs. For instance, for drug screening, SciOne AI integrated the customer's proprietary tools into the multi-agent system to create an end-to-end automated pipeline that covers data acquisition, analytical tool selection and execution, and results delivery. This significantly boosted task processing intelligence, demonstrating the architecture's scalability and adaptability in complex business scenarios.