Coding agents pair a large language model with a harness that plans, edits, and executes code. A new technical report on arXiv examines how to supervise such agents before they act, rather than only after failures occur.
The paper introduces Cortex, a supervisory layer aimed at pre-execution governance, requirement tracking, and iterative validation and repair. In practice, this means checking planned actions against requirements before code runs, and looping back for fixes when validation fails.
The report's title describes a fixed-point theory of governed coding agents, suggesting a formal model of this supervision loop. However, the abstract is brief and does not include implementation details or experimental results, so the practical impact of Cortex remains to be seen.