Hammer House
Six coordinated AI agents running in production.
Problem
A single large model prompted to do everything gets unreliable fast — the product needed distinct jobs (research, drafting, review, scheduling, and more) handled by specialized agents that hand off to each other cleanly.
Role
Designed the multi-agent architecture: how each of the six agents is scoped, how they hand off work to one another, and how the system stays observable instead of becoming a black box.
Stack
- Next.js
- AI Models
- Agent Orchestration
- Supabase
Architecture
Each agent owns one job and one handoff contract. A coordinator routes work between them and persists intermediate state, so a failure in one agent doesn't take down the whole chain.
- Step 1
Intake agent
Classifies the incoming task and routes it to the right specialist agent.
- Step 2
Specialist agents (×6)
Each handles one scoped job — research, drafting, review, and more — with a defined output contract.
- Step 3
Coordinator
Sequences handoffs between agents and persists state after each step.
- Step 4
Human checkpoint
Key outputs surface for review before anything ships downstream.
Outcome
Six agents with clearly scoped responsibilities operate as one coordinated system in production, rather than one oversized prompt trying to do everything at once.