Safety-Constrained Agentic AI for Autism Screening: A Multimodal, Clinician-Guided Architecture

Checked 30 Sept 2026. Open license (CC-BY, CC-BY-SA, CC0, public domain): full text indexed and used in synthesis.

Autism spectrum disorder (ASD) diagnosis often encounters substantial delays due to several reasons, such as shortages of trained specialists and limited access to care in rural and underserved communities. Moreover, it is very difficult to perform behavioral assessments within a single clinical visit, as it is significantly dependent on the child's behavior. Delayed diagnosis can postpone early intervention, which is important for improving developmental outcomes in children with ASD. Although artificial intelligence (AI) is increasingly explored in healthcare, its adoption in ASD screening remains limited due to concerns about reliability, governance, consent management, bias, and clinical trust. In this work, we propose a conceptual, governance-driven, clinician-augmented AI framework designed to assist clinicians during the ASD screening process rather than replace them. The proposed architecture collects various inputs such as text, audio, and video of a child from parents, schools, or caregivers, and then it runs through multiple specialized agents who are responsible for consent validation, bias monitoring, model selection, confidence-based abstention, and providing a structured report which will help clinicians in their assessment. Caregivers receive only non-diagnostic guidance, while clinicians receive structured decision-support information designed to aid clinical evaluation. The main goal of this article is not to validate model performance. We are mainly trying to design an agentic framework where governance and safety rules can be managed properly through multiple specialized agents. Although we have performed a single model training using a ResNet-50 facial-image classification model on a publicly available dataset, our main goal was to validate the governance and multi-agent system. The Stage 1 governance validation was done using more than a hundred scenarios. It is very important to highlight that our article should be viewed as a conceptual governance-driven agentic framework with Stage 1 validation, and it is definitely not a fully workable clinical solution. In the next phases, we plan to collect clinically validated data and focus more on model training, multimodal integration, and real-world validation.

• Patra et al. 2026 build governance directly into their autism-screening architecture: an orchestration layer coordinates specialized agents for consent verification, bias and applicability checks, and confidence evaluation, deliberately kept as separate responsibilities so that every decision remains transparent and auditable. Trace →
• Patra et al. 2026 acknowledge that their system is not yet production-validated: its abstention and governance behavior was evaluated only in a fully controlled environment, and how the system would integrate into existing clinical workflows remains unevaluated. Trace →
• Eranga et al. 2025 and Patra et al. 2026 independently converge on separation of concerns as the foundation of auditable production agentic systems: Eranga prescribes single-responsibility agents and clean separation between workflow logic and MCP servers, while Patra assigns each architectural layer a clear responsibility so governance checks stay visible and testable. Trace →
• The evaluation gap that Mehta 2025 quantifies in benchmarks—missing multidimensional metrics and inadequate reliability assessment—is illustrated concretely by Patra et al. 2026, whose governance and abstention behavior was validated only in a controlled environment rather than under real-world deployment conditions. Trace →

• Production agentic workflows and evaluation