HiveMind

The Problem

There was no registry of the specialised assistants I'd built, no memory that outlived a session, and no way to ask one question and have it land with the right agent. I wanted a personal orchestration layer. One endpoint, many agents, state that persists.

What It Does

You register agents. Each has a model, a system prompt, a temperature, a job. Then you query the hive rather than any individual agent, and it decides who answers. If the request reads like several tasks stitched together, it writes a delegation plan instead, fans the subtasks out in parallel and compiles a single reply. Agents can be pushed into structured collaboration, whether parallel, sequential, debate, consensus, or hierarchical, with roles assigned per mode. Each one carries its own vector memory, so it recalls earlier conversations. It's reachable two ways: as an MCP server over stdio inside Claude Desktop, and as a REST API behind a React chat interface.

The hard part

Routing a question to the right agent is the easy half. The hard half is memory: what each agent is allowed to recall, and what that recall costs to keep running. Give every agent a proper memory service of its own and you have bought an operations problem for a system with exactly one user, something to host, watch and pay for long before the idea has proved anything. Give them none and they are just prompts wearing names.

The second decision was more ambitious and I am less confident it was right. I let the system stand up its own hosting, so an agent could provision what it needed and remove it afterwards. It is a TypeScript system and that part does work. It is also the least tested thing I have written, which is an uncomfortable pairing.

Where It Stands

The whole thing lived inside a single month of 2025, in a few concentrated bursts. It stopped where I stopped, and nothing has been added since.

Stopped, deliberately. HiveMind does what it was built to do: register agents, route a question to the right one, and run several together on a task that needs more than one. It is an experiment I consider finished rather than a product I am maintaining. It taught me how agent systems hold together, and that understanding went straight into the work that followed. I would rather call it done than keep a one-user system on life support.