Agent harnesses let large language models act through tools, observe outcomes, and iteratively pursue a task goal. Verification checks and revises the model's output to improve quality, yet systematic guidance on organizing verification is scarce. We treat verification as a test‑time scaling axis, examining when to intervene, who performs the check, and how checking leads to revision. Our framework defines four layers: initial instructions, mid‑turn reminders, completion gates, and independent review. We evaluate 17 designs across seven frontier coding models on Terminal‑Bench (TB) and SWE‑bench Verified (SWE), and introduce VisualBench for expert assessment of visual and interactive artifacts. Results show that initial instructions and execution‑time reminders yield modest, model‑dependent gains with negligible extra output tokens. Completion gates can produce larger gains but at a substantially higher cost. Independent review by a critic offers a more favorable accuracy‑cost trade‑off, improving TB accuracy by 5.9‑9.5 percentage points and SWE by 3.9‑6.6 points, while incurring only 1.13‑1.25× and 1.00‑1.04× the baseline cost respectively. On VisualBench, independent review achieves net preferences of 13.2‑23.8 points over the baseline across four designs, surpassing all L0‑L2 variants. Adding review rounds or extra independent critics further lifts TB average accuracy by up to 3.5 points. Overall, these empirical findings provide practical guidance for verification design in agent harnesses and endorse independent review as an effective test‑time scaling axis.
Review