Long-running agents quietly take shortcuts, so the workflow state moves out of context: Haiku 4.5 follows the procedure on 139 of 200 accounts
Language-model agents drift from a procedure stated once at the start of a long unattended run; rewriting that procedure as an executable finite-state machine enforced by the runtime at tool calls, state transitions and turn endings let Claude Haiku 4.5 average 139.2 faithful and correct decisions on a 200-account SOP-Bench batch, against 48.4 for the strongest baseline, Remind-loop. Restating the instructions or resuming a stopped agent only moves the drift to another shortcut, and only closing all three routes held the procedure to the last account.