ATC Guardian - Multi Agent Support for ATC

ATC Guardian is a cross-framework, multi-agent decision-support system for Air Traffic Control, built for the Band of Agents Hackathon (Track 3: Regulated and High-Stakes Workflows). Six specialized AI agents — built on LangGraph, Pydantic AI, and CrewAI — collaborate through Band to detect aircraft conflicts, analyze hazardous weather, and coordinate emergency responses in real time. The problem: ATC is a regulated, safety-critical domain where every decision must be auditable. Today, AI agents operate in isolation — one framework, one workflow, no cross-examination. But real ATC work is collaborative: controllers, weather desks, and emergency responders constantly share context, cross-check each other, and escalate to human authority. Our solution uses Band as the genuine collaboration layer. When the system detects a converging conflict, it @mentions the Conflict Detector, which computes closest-point-of-approach and issues a structured advisory. That advisory is @mentioned to an independent, adversarial Safety Reviewer that re-derives the math against ICAO separation minima and returns APPROVE, REJECT, or MODIFY. Only then does the Coordinator queue the decision for the human controller — who holds sole authority to execute. Nothing an agent recommends is actioned without a human click. The Emergency Response agent recruits Ground Ops into the cascade for runway information on squawk 7700, and holds veto power deferring lower-priority advisories per ATC priority rules. Every thought, tool call, verdict, and controller resolution is written to a regulator-ready audit log exportable as JSON. A unique differentiator: controllers can propose a maneuver and preview the predicted CPA outcome before acting. Target audience: air navigation service providers, airline operations centers, and any regulated domain where review, traceability, and careful decision-making matter.