Overview
When multiple AI agents share a goal without coordination rules, their failures reveal the hidden operating system of the task.
The research explores what happens when you deploy multiple AI agents on shared tasks without giving them explicit rules for coordination. Do they develop emergent protocols? Do they compete or cooperate?
Key Questions
- How do agents negotiate task boundaries without explicit instructions?
- What patterns emerge when agents have overlapping capabilities?
- Can we observe something like "culture" forming in agent collectives?
- How does swarm behavior change with different LLM backends?
Methodology
Run matched agent groups through shared tasks, vary coordination rules and model backends, then score collisions, handoffs, loops, and recovery.
Tools & Frameworks
- Multi-agent task runs with explicit and implicit coordination rules
- Model and prompt variations to expose brittle handoff behavior
- Review traces that separate useful delegation from duplicated work
Preliminary Findings
Early signal: agents do not just fail independently. They create coordination debt: duplicated work, silent handoffs, and confident collisions.
"The hard part is not getting agents to act. It is getting them to notice what the other agents already changed."
What's Next
Next up: tighter task traces, clearer handoff scoring, and more examples where coordination pressure changes the result.