Building a Multi-Agent Air Traffic Control System

MaryamMiradi · x · 2026-07-18

Based on relevant research, the author designed and replicated an Air Traffic Control (ATC) system featuring 6 collaborating AI agents to manage safety in complex airspace (e.g., scheduling 30 aircraft simultaneously).

Division of Labor Among 6 Core Agents:

Agent 1 (Itinerary Planning): Generates complete flight routes for each aircraft, issuing instructions based on relative distance rather than absolute time to adapt to dynamic factors like wind speed.
Agent 2 (Airspace Simulator): Runs a digital twin every few seconds to predict conditions up to 60 minutes ahead, conducting over 20 stress tests under extreme scenarios (e.g., strong winds, radio silence).
Agent 3 (Hazard Monitoring): Monitors all aircraft pairs, categorizes conflicts (e.g., head-on approaches, overtaking), and matches them with appropriate solutions.
Agent 4 (Problem Solver): Attempts fixes by priority, re-running simulations after each adjustment. It backtracks if new issues arise, ensuring global safety.
Agent 5 (Root Cause Analysis): Pinpoints the specific instructions causing hazards, making only minimal modifications to avoid contradictory commands.
Agent 6 (Execution & Safety Net): Dispatches instructions at precise moments; if all automated fixes fail, it moves aircraft to safe altitudes, hands control over to humans, and clearly explains "why this action was taken."

Core Design Philosophy: Simulate before executing, fix root causes rather than surface symptoms, and maintain Human-in-the-loop intervention when system limits are reached.

Original post →

Building a Multi-Agent Air Traffic Control System

More from coding & agent