Substrate-Aware AI Agents: Execution Context as a First-Class Input

Recent research indicates that if artificial intelligence agents are not aware of their operating environment (i.e., “submatrix blindness”), it will severely impair their execution capabilities. The experiment compared the performance of three models: Anthropic Claude Opus 5, OpenAI GPT-5.6-Sol, and Google Gemini 3.7 Flash, under two conditions: when only the task description was provided, and when an execution context containing 128 MB of RAM and a wall time constraint of 10.0 seconds was given. The results showed that disclosing the execution contract reduced peak process memory, decreased average wall time, and increased execution speed by up to 3.1 times. Under more stringent resource constraints of 96 MB, the correct and within-budget results for the three models were 4/5, 5/5, and 3/5, respectively, while the accuracy under the task-only mode was 0/5 or 1/5.