Quitting Is a Setting: An Unseen Reasoning Tier Makes a Frontier Agent Abandon a Quarter of Its Tasks

A frontier coding agent abandons a quarter of its tasks. It reads the problem, runs commands, and then voluntarily stops, having solved nothing. The only thing that has to change to prevent this is a reasoning-tier setting that never appears in its prompt. We give agents six independent, deterministically graded bugs in a real repository, set no budget, and vary only the tier: gemini-3.1-pro abandons 14 of 56 episodes at LOW, 0 of 56 at MEDIUM and 0 of 56 at HIGH (Fisher $p = 6.2 \times 10^{-8}$). Abandonment is not exhaustion, since those episodes stop after a median 4 steps and spend only 3–4% of the configuration's own unconstrained consumption. Every one of them ran the test suite and saw it fail. Capability is present at LOW, where 23 of its 56 episodes still score a perfect 6/6. We then tell the agent not to stop, in one arm by saying its setting does not limit it and in another by forbidding it to finish before attempting every bug. We still see 13 of 78 abandon ($p = 5.6 \times 10^{-4}$ against the MEDIUM control). In the latter, 27 of 38 episodes demonstrably read the instruction. Asked why they stopped, four of the fourteen cite an effort setting they were never shown. We release the harness, the repositories, and every trace (Appendix B).