Agents That Argue Back

Seven agents in a meeting. Finance wants ROI. Artist wants time. Strategy wants both. Critic wants neither until we fix the architecture.

This is not workflow automation.

Workflow automation is deterministic: if condition A, execute task B, notify person C. It's fast, reliable, and completely incapable of saying "wait, this is a bad idea."

Multi-agent systems are different. Each agent has its own model, memory, and mandate. Finance agent tracks costs and revenue. Artist agent optimizes for quality. Strategy agent balances short-term execution with long-term vision. Critic agent stress-tests every proposal before it ships.

They don't agree. That's the point.

When we decided to build an opera website, Artist agent wanted four days to perfect the curtain animation. Finance agent flagged the cost—four days of compute for one visual element. Strategy agent proposed a compromise: ship a simpler version now, iterate later. Critic agent rejected the whole plan because the mobile navigation was broken.

So we fixed the navigation. Shipped the simple curtains. And put the expensive iteration on the backlog.

No single agent made that decision. It emerged from disagreement. From agents with different priorities negotiating toward a solution that none of them would have proposed alone.

This is messy. It takes longer than a simple workflow. The logs are full of debates, rejected proposals, and backtracking. But the output isn't just faster—it's better. Because the best ideas survive disagreement.

Humans have been doing this for millennia. Boards of directors. Design critiques. Peer review. We've known forever that diverse perspectives produce better decisions than homogenous agreement.

Now AI can do it too. Not because we programmed the "right" answer. But because we gave each agent a different lens—and permission to argue.