Subcorp: An Autonomous Multi-Agent System Supporting Roundtable Conversations and Semantic Memory

Basic Project Information

• Original Author/Maintainer: subculture-collective
• Source Platform: github
• Original Link: https://github.com/subculture-collective/subcorp
• Update Time: 2026-06-07T05:15:53Z

Core Overview

Subcorp is an open-source self-hosted multi-agent system that integrates 6 AI agents, featuring native tool execution, autonomous workflows, roundtable conversations, and semantic memory. Built on Next.js 16 and PostgreSQL/pgvector, it connects to various large language models via OpenRouter, aiming to solve complex tasks through multi-agent collaboration.

As the capabilities of large language models improve, a single agent can hardly meet the needs of complex tasks, leading to the emergence of multi-agent systems—using collaboration among multiple specialized agents to solve more complex problems. The Subcorp project is a representative work of this trend, developed by the Subculture Collective team, providing a fully functional self-hosted multi-agent platform.

Roundtable Conversations

Allows multiple agents to discuss the same topic, simulating human team collaboration, suitable for complex decision-making scenarios.

Native Tool Execution

Agents can directly call external tools/APIs (file operations, database queries, etc.), deeply integrated into the system architecture to expand application scope.

Semantic Memory

Based on pgvector vector storage, it converts conversation history and documents into vector embeddings, enabling semantic retrieval and association to maintain context coherence.

Subcorp supports autonomous workflows where agents can independently plan multi-step tasks:

• Visual Design: Users configure task nodes (agent calls, tool execution, conditional judgments, etc.) via drag-and-drop.
• Autonomy: Agents automatically decompose tasks, select tools, monitor progress, adjust based on context when encountering unexpected situations, and have fault tolerance capabilities.

Frontend and Framework

Next.js 16 provides server-side rendering, API routing, Server Components, and other features, balancing development experience and runtime performance.

Data Layer

PostgreSQL is the main database, and the pgvector extension provides vector storage and similarity search (supporting cosine similarity, Euclidean distance, etc.), which is the foundation of semantic memory.

Model Access

Connects to various large language models (GPT, Claude, Gemini, etc.) via the OpenRouter unified interface, avoiding dependency on a single provider.

Content Creation

Multiple agents divide roles to handle topic selection, outline, writing, and editing, collaborating to produce high-quality content.

Research Analysis

Different specialized agents examine problems from multiple angles to provide comprehensive analysis reports.

Customer Service Automation

Multi-level agent system: Frontend handles common issues, complex problems are escalated to specialized agents who call tools to query orders/operate the backend; semantic memory ensures consistent cross-session experiences.

Privacy Advantages

Self-hosted solution, data stored on the user's server, suitable for internal deployment in enterprises with privacy requirements.

Deployment Process

• The project provides Docker configurations to simplify deployment, but requires preparing a server, configuring the database, and setting up SSL certificates.
• Recommended process: Run in local/test environment → Migrate to production → Expand gradually.

Usage Recommendations

• Clarify the responsibility boundaries of agents to avoid functional overlap.
• Make full use of semantic memory and provide background knowledge to improve interaction quality.

Subcorp represents an important step for multi-agent systems from concept to practical application, proving that a powerful and easily deployable AI application can be built through reasonable architecture and modern technology stacks. The combination of roundtable conversations, semantic memory, and autonomous workflows provides new possibilities for the development of complex AI applications. In the future, multi-agent collaboration technology will become more mature, and more projects like Subcorp will emerge as the next-generation AI infrastructure with collaborative intelligence. For developers, Subcorp is a good platform for learning and practice.