I moved the agent topology out of my code and into the database

说明与用途

将代理拓扑从代码迁移到数据库，减少了部署开销，使领域专家能够通过管理面板编辑代理行为，并引入多代理工作流，在不更改代码的情况下捕获错误。

问题

Developers spent a month building invisible infrastructure—agent loops, tool registries, data storage, and workflow rules—each change requiring a deploy, making maintenance costly and opaque.

结果

Moving agent topology to the database reduced deployment overhead, enabled non‑technical domain experts to edit agent behavior, and introduced multi‑agent workflows that catch errors without code changes.

工作流

• 1. Define tools as OpenAI function specs with editable Python bodies in the tools table.
• 2. Create skills that bundle instructions and required tools, stored in skills and skill_tool tables.
• 3. Assemble agents with system prompts and associated skills, stored in agents and agent_skill tables.
• 4. Design workflow graphs (start/group/stop) in projects.workflow JSON to control agent order and delegation.
• 5. Store domain data and schema descriptions in real tables; column descriptions serve as prompts for tools.
• 6. Use a context tool to run pre‑message logic, stored in agent_context_tools.
• 7. Run agents via a single uvicorn worker; tool bodies execute with full builtins and are admin‑reviewed.
• 8. Admins edit tool bodies or agent prompts through the panel; changes cache by body hash and apply to next message.

证据

I built it once. I open‑sourced it as CerebrumKit.

来源

Editing a tool body in the panel takes effect on the next message.

来源

Tool bodies are code execution. They run with full builtins and are not sandboxed.

来源