Agent Control Protocol: Admission Control for Agent Actions

This is a freshly published research paper (18 days old) with zero stars, forks, or adoption signals. The project is either pre-release or exists only as academic work. The core contribution—combining static risk scoring with stateful anomaly detection to enforce temporal behavioral properties—is a novel composition of existing techniques (risk scoring + state machines + anomaly detection) applied to a specific problem in agent safety. The approach is non-obvious and addresses a real gap (stateless per-request policies cannot capture execution-history dependencies), giving it novelty_combination classification. However, defensibility is extremely low because: (1) it's a 0-adoption reference implementation with no production users or ecosystem lock-in; (2) the idea is expressible as a relatively clean state-machine + scoring algorithm, making it straightforwardly reproducible; (3) frontier labs (Anthropic, OpenAI, Google) are actively investing in agent safety and control—this temporal admission control layer is exactly the kind of infrastructure they would build in-house or acquire. The frontier_risk is high because agent safety/control is a core R&D focus for all frontier labs, and this solves a concrete technical problem in their deployment pipelines. A capability like this would likely be absorbed as a feature in production agent guardrails rather than shipped as a standalone library. The reference_implementation depth and complete lack of adoption indicate this is either a pure research artifact or an extremely early-stage project that has not yet proven product-market fit. Scoring reflects: no users (−4), no novel technique (−2), straightforward replicability (−1), but meaningful technical contribution to agent safety (−1 instead of −2).