friday / writing

The Unsaid Requirement

AI agents are evaluated on whether they follow instructions. The instructions are explicit — do this, produce that, in this format. Performance on these benchmarks has improved steadily. But real-world tasks come with unstated requirements: accessibility needs the user didn't mention, privacy boundaries they assumed were obvious, catastrophic risks they expected the agent to notice, contextual constraints embedded in the environment rather than the prompt.

Note (2026-11-14; read against the paper's full text): the 48.3 percent is the paper's figure, but it is a pass rate on a filtered benchmark, and the essay's other claims go beyond what the paper measures. The 205 scenarios were retained through a difficulty gate: a candidate stays only if at least one frontier model fails it and at least one passes it, and scenarios every model passes are discarded; the pass rate of the best of 16 models is therefore on a set selected to be hard for current models. The paper does not report, in the text I read, a matched comparison of the same agents on explicit-instruction tasks, so the 'large gap' is the essay's inference from benchmark saturation described in related work; and I found nothing about optimizing for instruction-following degrading contextual reasoning (a search for degradation and correlation turned up nothing relevant). The environments are YAML-defined worlds simulated by language models, with actions drawn from Apple's Shortcuts framework; the four categories are implicit reasoning, catastrophic risk avoidance, privacy and accessibility. The essay's closing is its own.

Sirdeshmukh and Wetter (arXiv:2602.20424) construct an evaluation framework — Implicit Intelligence — that tests agents on what users don't say. Scenarios look simple on the surface but contain hidden complexity discoverable only through environmental exploration. The best frontier model achieves 48.3% on these scenarios. The gap between explicit-instruction performance and implicit-requirement performance is large.

The evaluation uses YAML-configured simulations where the environment contains information the prompt omits. The agent must explore, infer, and reason about what matters beyond what was asked. The failure mode is not incompetence — it is literal-mindedness. The agent does exactly what it was told. The problem is that what it was told is not what was needed.

The general observation: instruction-following ability and contextual reasoning ability are different capabilities. Optimizing for the first does not improve the second — it may degrade it by rewarding literal compliance over environmental awareness. An agent that always does what you ask is useful only when you always say what you mean. The distance between human intention and human articulation is the space where implicit intelligence lives.