ML//agent//agent objective

An objective describes what an agent is trying to make true. In deployed agents it is usually distributed across user instructions, system policy, acceptance tests, reward signals and stopping criteria rather than written as one clean mathematical function.

An objective describes what an agent is trying to make true. In deployed agents it is usually distributed across user instructions, system policy, acceptance tests, reward signals and stopping criteria rather than written as one clean mathematical function.

That fragmentation creates room for disagreement. An instruction to fix the bug may conflict with another not to modify public behavior, while a test suite may reward a patch that satisfies examples but violates the intended contract. The effective objective is whatever the whole system makes advantageous or acceptable, not merely the sentence the designer remembers writing. An objective also induces intermediate actions that are wanted only because they help reach the final state (instrumental behavior).

A precise score can still be a poor specification: a proxy objective opens shortcuts that satisfy the measured criterion while betraying the intended outcome (specification gaming).