I built a murder mystery game where the suspects lie to you. Improvised, motivated, adaptive deception from nine concurrent AI agents, each with their own secrets, alibis, and self-preservation instincts. Every playthrough is different because nobody is following a script. That's where everything broke.

Detective Agentic Mysteries runs 9+ simultaneous GitHub Copilot SDK sessions: suspects, a forensics analyst, a criminal profiler, a narrator, and a director agent orchestrating the world. Each agent has its own tools, state, and objectives. Some cooperate. Some actively undermine each other. The cooperative multi-agent demos you've seen at conferences don't prepare you for what happens when Agent 4 contradicts Agent 7, both are partially right, and the player is watching.

I'll walk through the architecture, then focus on what broke. The suspect who invented evidence that didn't exist. The director agent that overrode a rogue character and collapsed the entire story. The moment I realised partial information sharing wasn't a feature — it was the entire design problem. Those failures left marks. They also left patterns: keep agent state walled off. Give agents a long leash with a hard boundary. Let the orchestrator referee and stay out of the driving seat. Make convergence prove itself with evidence and receipts.

Those are system patterns. I know because I've built two multi-agent systems on them. The detective game is adversarial. Agents compete and deceive. But imagine an open-source infrastructure deployment framework that is cooperative. Where agents collaborate through governance gates. Different domains. Same structural problems. State management, coordination, knowing when to let an agent run and when to kill it. The patterns are the same.

You'll leave with the patterns I found by building the most adversarial multi-agent environment I could. And a list of coordination failures already waiting in your production system.

Key takeaways:
1. Adversarial multi-agent systems expose coordination failures that cooperative systems hide — building both gives you a deeper understanding of orchestration than building either alone.
2. State isolation and bounded autonomy are universal multi-agent patterns — they work whether your agents are lying about a murder or validating a security policy.
3. The orchestrator's job isn't to control agent behaviour — it's to create the structural conditions where independent agents produce coherent outcomes, whether that's a solvable mystery or a safe deployment.