TT-Symbiote: A Shared Knowledge Base Architecture for Agent Workflows

This article introduces the TT-Symbiote shared knowledge base architecture, which aims to address collaboration challenges in multi-agent workflows. Serving as a cache layer between the Team Prompt Flow and Research Prompt Flow, this architecture enables knowledge sharing and state management among agents. Its core values include decoupling agent collaboration, enhancing system observability, and improving dynamic scalability.

With the advancement of LLM capabilities, multi-agent workflows have become an important paradigm for AI application development, but they face core issues: difficulty in efficiently sharing knowledge and state between different agents. Traditional tightly coupled designs are prone to single points of failure, hard to scale and maintain, and complex to debug. The TT-Symbiote project emerged to solve this problem by using a shared knowledge base as an intermediate layer.

System Roles

• Team Prompt Flow: Led by the Architect Agent, responsible for task planning and team coordination (deciding "what to do" and "who will do it").
• Research Prompt Flow: Executed by Guide Generation Agents, responsible for specific research tasks and knowledge generation (focusing on "how to do it" and concrete outputs).

Intermediate Layer Role of the Knowledge Base

Undertakes responsibilities of state caching, knowledge transfer, result aggregation, and collaboration synchronization.

Core Design Philosophy

• Decoupling and Caching: Introduce an intermediate layer to decouple agents; each agent interacts only with the knowledge base.
• Persistence and Observability: Make workflow states observable and auditable, supporting breakpoint resumption and fault tolerance recovery.
• Dynamic Collaboration: New agents can join by following the access protocol without modifying existing code.

Data Model Design

Supports types like Topics, Guides, State, and Metadata.

Access Interfaces

Provides standardized CRUD interfaces, supporting reading task context, writing research results, state change subscriptions, and complex conditional queries.

Concurrency Control

Handles read-write conflicts and transactional updates to ensure consistency, choosing optimistic or pessimistic locking strategies.

Complex Research Tasks

Break down large tasks into subtasks, track progress, and integrate multi-agent results.

Team Collaboration Simulation

Simulate human team collaboration: Architect (project manager) makes plans, Guide Generation (researchers) execute research, and the knowledge base serves as a shared document and progress system.

Continuous Learning System

Accumulate historical results to form organizational memory, avoid duplicate research, and use past experience to guide new tasks.

Evolution of Agent Architecture

Represents the evolution direction from chain calls to collaborative networks, then to distributed architectures based on shared knowledge bases—similar to the development of human organizations from individual heroism to knowledge management organizations.

Value of Reusable Components

Independent knowledge base components bring reusability (applicable to different agent systems), testability (independent testing of read-write logic), and replaceability (replace underlying storage).

Implications for Community Collaboration

Uses GitHub repositories as carriers, leveraging Git version control and collaboration mechanisms to manage knowledge evolution.

For Agent Developers

• Introduce a shared knowledge base as an intermediate layer
• Design clear data models and access protocols
• Implement comprehensive logging and monitoring

For Research Teams

Apply to:

• Distributed research for literature reviews
• Decision support for multi-expert consultations
• Knowledge integration for interdisciplinary projects

Although TT-Symbiote is a niche project, it touches on core issues in agent system design: effective collaboration and knowledge sharing between distributed intelligences. As AI agent technology matures, such infrastructure will become increasingly important, providing a reference architectural pattern for building complex agent applications.