Autonomous AI agents must factor memory, execution time, compute budget, and operational limits when selecting actions; together these define a feasible plan. The omission of such execution context from an agent's planning state is termed substrate blindness. We evaluate this hypothesis via numerical code generation: for a high‑dimensional pairwise Euclidean‑distance task, three frontier models (Anthropic Claude Opus 5, OpenAI GPT‑5.6‑Sol, Google Gemini 3.7 Flash) generate code either from the task description alone or with a 128 MB RAM and 10 s wall‑time contract.
In 13 of 14 task‑only vs. contract‑disclosed comparisons, peak memory usage dropped, and all three cohorts showed reduced mean wall time, achieving up to 3.1× speed‑up. Disclosure of the contract induced structural code changes such as bounded blocking, float32 retention, upper‑triangle traversal, and in‑place or memory‑mapped buffers.
Under a tighter 96 MB contract, independently sampled contract‑disclosed cohorts achieved correct‑and‑within‑budget outcomes of 4/5 for Claude Opus 5, 5/5 for GPT‑5.6‑Sol, and 3/5 for Gemini 3.7 Flash, compared with task‑only results of 0/5, 1/5, and 0/5 respectively. Their mean MaxRSS and wall time were 49‑74% and 35‑64% lower than the task‑only baselines.
These findings constitute a controlled proof‑of‑concept: a minimal execution contract prompts proactive structural adaptation in generated programs, shifting computation away from unconstrained allocations and substantially improving observed resource‑time profiles before execution.
Review