This paper investigates how agent harnesses generate value in retail and airline use‑cases. We decompose an agent’s core functionality into three components:
- Planning information (provides task guidance);
- Execution organization (schedules subtasks and tracks state);
- Completion verification (uses a read‑only terminal to check results).\ \ Experiments were run on two retail setups and an airline pilot, all on the $\tau^2$ benchmark. To isolate the contribution of the planning content, we created two conditions:
- Fixed: hand‑crafted, task‑specific plans;
- Sham: policy text shuffled to match Fixed’s word count.\ \ Across 265 matched cells, Fixed improved oracle‑verified success by 7.17 percentage points (90% task‑clustered bootstrap interval 1.15–13.36), with gains concentrated in higher‑complexity tasks.\ \ A read‑only terminal verifier rejected 61% of retail oracle‑invalid episodes while only withholding 17% of correct ones, at less than one cent of additional cost per episode.\ \ The key insight is that the relative importance of components depends on the loss assigned to erroneous acceptance:
- With low liability, the planning gain dominates;
- With high liability, the verifier’s avoidance of false passes dominates;\
- A standalone verifier captures nearly all the false‑pass benefit of the full planning‑plus‑verification stack at a fraction of the cost.\ \ Review: The study highlights a practical cost‑benefit trade‑off between sophisticated planning and lightweight verification in stateful LLM agents. For high‑risk applications, a simple read‑only verifier may deliver more value than elaborate task‑specific plans.