An agent that builds agents

Led the second version of aiXplain's agent-building agent. Users describe what they need in a conversation, and the system routes it to agents for search, agent-building, debugging or chat.

Problem

People come to the platform with a request in plain language. Sometimes they want to find something, sometimes to build a new agent, sometimes to fix one, and sometimes they just have a question. The system had to work out which it was and send the conversation to the right place.

What made it hard

The main difficulty was routing. A message like "actually, make it faster" only makes sense next to the messages before it, so the classifier has to read the whole conversation, not just the last message.

The second challenge was knowing when to ask for clarification instead of guessing. If a user says "make an agent that helps with documentation", should I assume they want it to write docs, search them or review them? A wrong guess wastes the user's time.

What I built

I led the second version. A classifier reads the whole conversation and routes it to one of four agents: search, agent-building, debugging or chat. When it cannot tell what the user wants, it asks a clarifying question instead of guessing.

It runs as a Kubernetes service (FastAPI, Celery, Redis). Its tools are exposed as REST endpoints and as an MCP server, and I also built a terminal client for it.

What it meant

Building or fixing an agent can start from a conversation. The same tools are open to other agents over MCP and to programs over REST.

It is one of the three pieces of work aiXplain's CEO cited when he recommended me in 2026 for initiative and technical leadership.