A new multi-agent tool called HATS wants to replace the single-chatbot habit with something messier and more useful: a roomful of AI agents that openly disagree. According to Hacker News, where the launch climbed to 160 points, HATS borrows from Edward de Bono’s Six Thinking Hats framework and turns each hat into a separate AI agent with its own job, voice, and perspective. The pitch is direct. One AI tells you what you want to hear. Six AIs, forced into structured conflict, find the holes you missed.
What stands out here is the design choice to engineer disagreement into the system itself. LLMs love to agree with the user, sound confident when they’re wrong, and skip past blind spots. HATS pushes back by assigning roles that have to argue.
The Six Agents
- White Hat handles facts, gaps, and stated assumptions.
- Red Hat speaks for human impact and gut intuition.
- Black Hat plays critic, surfacing risks and failure modes.
- Yellow Hat argues the upside and opportunity.
- Green Hat brings creativity and alternatives.
- Blue Hat facilitates and synthesizes the room.
Each agent can be powered by a different model. The UI lets you mix OpenAI, Claude, and Gemini per role, or run everything locally through Ollama or LM Studio. Token use and cost are tracked agent by agent.
Live Meetings With Faces and Voices
This is where HATS gets ambitious. When a meeting starts, the app opens a stage of animated 3D avatars rendered in real time with Three.js. Each agent speaks aloud through Piper TTS using a per-agent voice model. Lip sync runs through Rhubarb, which maps phonetic timing to ARKit visemes so mouths actually move with the audio.
Five meeting types ship out of the box: Standup, Sprint Planning, Retrospective, Review, and Ad Hoc. Humans take turns in the conversation alongside the agents, and the full transcript exports as Markdown when the meeting ends.
A Kanban Board That Moves Itself
HATS includes a six-column board (Backlog, Ready, In Progress, Blocked, Review, Done) with drag-and-drop. The twist: when a ticket moves to In Progress, it gets dispatched to the assigned agent as a live task. Resolve a blocker and dependent tickets unblock automatically. Tickets aimed at the human lead glow orange so nothing slips.
Agents talk to the board through a Kanban MCP server, creating, moving, and commenting as part of their normal workflow.
MCP Integrations Out of the Box
Through the Model Context Protocol, HATS ships connectors across five categories:
- Productivity: Kanban, Memory knowledge graph, Slack
- Files and Documents: Filesystem, Excel, Word, PDF, PowerPoint
- Web: Brave Search, Puppeteer/Chrome
- Databases: SQLite, PostgreSQL
- Development: GitHub
Every server is togglable from the UI with live connection status.
Use Cases the Team Pitches
- Run a product planning meeting with structured pushback
- Stress-test a startup idea before you waste a quarter on it
- Explore trade-offs in architecture decisions
- Replace async brainstorming threads that go nowhere
Setup and Caveats
HATS runs on Node.js 20+ with a TypeScript backend and a plain HTML/CSS/JS frontend that needs no build step. Setup scripts cover Windows, Linux, and macOS. You need at least one LLM provider key (Anthropic, OpenAI, or Gemini) in the .env file. Brave Search is optional for the web tool.
Voice and avatars are optional too. Piper TTS runs in either a recommended server mode (one voice per agent, models held in memory) or a slower subprocess fallback that adds about half a second of latency per sentence. Skip Piper entirely and the agents respond in text.
Why this matters: most “AI team” products are thin wrappers around one model with different prompts. HATS commits to real multi-agent orchestration with persistent state, project isolation, and a workflow surface (the Kanban board) that the agents actually act on. Whether structured disagreement produces sharper decisions or just louder ones is the open question, but the architecture is a serious answer to the sycophancy problem. Full setup details and the demo video are at the original Hacker News post.