Originally published at thoughts.jock.pl
Anthropic dropped Opus 4.6 with Agent Teams capability. Instead of reading the release notes, I ran an experiment.
Four agents. Two projects. Running simultaneously.
The Setup
I told the agents to build two things in parallel:
Agent Orchestra -- an interactive visualization showing how AI agents coordinate with each other. Task delegation, message passing, real-time work completion with animations.
Dungeon of Opus -- a fully functional roguelike game. Procedural generation, turn-based combat, inventory, multiple enemy types with distinct AI behaviors.
No other instructions. Go build.
What Happened in 45 Minutes
Both projects shipped.
The game alone: 1,400 lines of functional code. Line-of-sight fog of war. Seven enemy types, each with unique behavior. BSP-style procedural level generation. Full inventory system.
Forty-five minutes.
What I Actually Learned
The impressive part is not the speed. It is what the agents figured out on their own.
Agent teams work best when tasks are genuinely independent. The moment dependencies emerge between parallel tasks, things get weird. Agents make assumptions about what the other is building instead of coordinating.
The other thing: creative direction is still firmly human territory. The agents executed. They did not decide what to build or why. That part was mine.
The gap between "I have an idea" and "there is a deployed product" got a lot smaller this week. But the gap between "I have an idea" and "that idea was worth building" did not change at all.
Good taste still matters. Maybe more now.
Try Both Projects
Both are live on wiz.jock.pl/experiments.
Want to see how I structure multi-agent systems? The AI Agent Blueprint covers the coordination patterns I use.
Top comments (6)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.