Multi-agent strategy design
Roundtable
A multi-agent system where three AI personas debate and refine strategy — from high-level vision down to task-level execution plans.
The Problem
Ask a single AI for a strategy and you get one perspective, delivered with the same confidence whether it’s right or not. Nothing pushes back, nothing stress-tests the reasoning, and nothing surfaces the tradeoff you didn’t think to ask about. That’s fine for a quick opinion. It’s not enough for a decision worth getting right.
Roundtable puts that pressure back in before the decision gets made instead of after — a panel of AI personas that has to work through disagreement over multiple rounds before it can call a result final.
The Three Personas
- Domain Expert — Grounds discussions in practical reality and established best practices
- Creative Problem Solver — Pushes boundaries with novel approaches and unconventional solutions
- Critical Analyst — The built-in contrarian. Pressure-tests every idea before it makes it into the final plan, killing groupthink
Getting these personalities right was crucial — different enough to create productive tension, but aligned enough to work toward common goals.
Demo
Why It’s Interesting
- The system explicitly tags which tasks need a human and which an AI can handle — no hand-waving about “AI will do it”
- Output isn’t just a strategy doc. It decomposes all the way to tasks with deliverables, skills needed, success criteria, time estimates, and cost projections
- Agents genuinely build on each other’s insights rather than generating independent responses — this required careful prompt engineering and state management to maintain context throughout conversations
Key Design Decisions
Billed to a subscription, not metered by the call. A single deliberation is a long-running job — a 3-round, 4-agent collaboration is roughly 13 sequential Claude calls. Rather than run that against a pay-as-you-go API key, the app calls Claude through the Claude Agent SDK using my local Claude Code login, so usage is covered by the Pro/Max subscription instead of stacking up metered API cost.
The database is the source of truth; the stream is just presentation. Every completed turn is persisted as it happens. If the connection drops mid-deliberation, the live view disconnects but nothing is lost — the run is still there when you come back.
Editing an agent never rewrites history. Each session snapshots the exact panel configuration it ran with, so changing a persona’s instructions later doesn’t silently alter the transcript of a session that already happened.
Tech Stack
Next.js 16 (App Router), TypeScript, Tailwind CSS v4, the Claude Agent SDK for Claude access, libsql (SQLite-compatible) for persistence, and Zod for validation.
