I built a free multi-agent AI debate system — no API keys, no cost, runs in OpenCode

#productivity #tutorial

I got tired of asking one AI model and trusting it blindly.

So I built SYNAPSE — a workflow where 3 AI agents debate your problem independently, then I only decide where they disagree. The whole thing costs $0 to run.

Here's how it works and how you can use it today.

The Problem With Single-Model AI

When you ask Claude or GPT a question, you get one perspective. That model has blind spots. You don't know what it missed.

The research backs this up — multi-agent debate systems consistently outperform single models on complex problems (Google DeepMind published on this in 2023).

But setting up multi-agent workflows is painful. API keys, orchestration code, rate limits, costs...

The Solution: A Workflow, Not a Program

SYNAPSE isn't a CLI or an app. It's a workflow pattern that runs entirely inside OpenCode using its free built-in subagents.

The workflow:

1. Write your problem to problem.md
2. 3 agents debate it independently:
   - @explore  → code-level analysis
   - @general  → architecture & design
   - @research → best practices & alternatives
3. Main agent reads all 3 opinions → writes consensus.md
4. You see: what they agreed on + where they disagreed
5. You decide the conflicts (takes 30 seconds)
6. Execute

No API keys. No TypeScript code to maintain. No costs.

Real Example: Mobile Auth Bug

Here's a session I ran on a real bug:

Problem: JWT auth works on web but fails for mobile users with 401 errors — even with fresh tokens.

What each agent found:

@explore → JWT verify() has no clockTolerance + Redis TTL has no buffer
@general → Clock skew is the root cause, fix it at the middleware level
@research → clockTolerance: 300 is the standard fix + send server timestamp for mobile

Consensus (91%):

✅ All 3 agreed: add clockTolerance: 300 to JWT verify
✅ All 3 agreed: add notBefore: '-30s' to token signing
⚠️ Conflict: fix refresh endpoint now or later?

My decision: Later — mobile release in 3 days, minimum risk.

Result: 3 files changed, mobile auth fixed, web auth unchanged.

The key thing: @explore caught the Redis TTL issue that wasn't even in the original problem description. One model would have missed it.

The File Structure

synapse/
├── AGENTS.md          ← the agent reads this, you don't have to
├── prompts/
│   ├── explore.md     ← tuned for code analysis
│   ├── general.md     ← tuned for architecture
│   └── research.md    ← tuned for best practices
└── demo/              ← complete example session

The main agent (Big Pickle in OpenCode) reads AGENTS.md and knows exactly what to do. You just describe your problem.

How to Use It

# 1. Install OpenCode (free)
npm install -g opencode-ai

# 2. Clone/download SYNAPSE
cd synapse

# 3. Open OpenCode
opencode

# 4. Tell the agent:
Read AGENTS.md and run a SYNAPSE session on: [YOUR PROBLEM]

That's it.

What I Learned

Multi-agent debate works — not because the models are smarter together, but because they catch each other's blind spots before the answer reaches you.

The workflow pattern > the code — I started building a TypeScript CLI with API adapters. Then I realized: the value is in the prompts and the workflow convention, not the code. Deleted the code, kept the markdown files.

Human-in-the-loop is the key feature — the system doesn't auto-execute contested decisions. You decide conflicts. That 30-second decision step is what pushes accuracy from 78% to 94%.