In traditional control theory and multi-agent systems, the default assumption is periodicity. Sensors sample at fixed intervals, controllers recalculate trajectories every few milliseconds, and agents broadcast their status to teammates in a constant "heartbeat."
While simple to implement, this "always-on" approach is remarkably wasteful. In a stable environment, most of these updates carry zero new information. For a swarm of drones, this waste translates to drained batteries; for Large Language Model (LLM) agents, it manifests as astronomical token costs and sluggish response times.
A growing body of research suggests a more surgical alternative: Event-Triggered Coordination. Instead of acting on a clock, agents act only when "something interesting happens."
The Efficiency of Management by Exception
The core intuition behind event-triggered systems is "management by exception." As Solmaz S. Kia and Sonia Martı́nez demonstrate in Distributed convex optimization via continuous-time coordination algorithms with discrete-time communication, agents can solve global optimization problems by only broadcasting updates when the error between their current state and their last transmitted state exceeds a specific threshold.
This shift from time-driven to state-driven logic can be transformative. In the realm of pathfinding, Biplav Srivastava and colleagues introduced IO-MAPF Scalable Multi-Agent Path Finding using Collision-Aware Dynamic Alert Mask and a Hybrid Execution Strategy, which reduces inter-agent communication by up to 23x. By using a "collision-aware alert mask," the system only forces agents to communicate when a central monitor predicts a potential conflict, allowing them to operate independently the rest of the time.
Reactive vs. Proactive: Event vs. Self-Triggered
Researchers generally divide these efficient strategies into two categories:
- Event-Triggered (Reactive): The system constantly monitors a condition (like a safety margin or an error threshold). When the condition is met, it triggers a calculation. Andrew J. Taylor and Pio Ong apply this to safety-critical systems in Safety-Critical Event Triggered Control via Input-to-State Safe Barrier Functions, ensuring that control updates only happen when the system approaches the edge of its "safe set."
- Self-Triggered (Proactive): The agent uses a model of the world to predict when it will next need to act. Manuel Mazo and Adolfo Anta describe this in Self-Triggered Control: trading actuation for computation as a way to "trade actuation for computation." Instead of listening for a trigger, the controller calculates a "deadline" for its next update based on current dynamics, allowing the hardware to sleep in the meantime.
Romain Postoyan and Dragan Nešić show that these proactive strategies can reduce network traffic by roughly 80% Event-triggered and self-triggered stabilization of distributed networked control systems while maintaining the same stability guarantees as high-frequency periodic systems.
Scaling to LLM-Based Agents
The move toward event-driven logic is now jumping from classical robotics to the world of LLM agents. In these systems, the "resource" being saved isn't just battery power—it's the context window and the latency of expensive model calls.
In the CaPo framework Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation, Jie Liu and Pan Zhou move away from greedy, step-by-step LLM prompting. Instead, they use "progress-adaptive execution," where agents only re-evaluate their shared strategy when significant task milestones are reached or unexpected obstacles appear.
Similarly, the CONAGENTS framework by Zhengliang Shi and colleagues Learning to Use Tools via Cooperative and Interactive Agents uses an adaptive protocol that only triggers "Reviewer" agents when the "Execution" agent encounters an error. This "management by exception" allows smaller, open-source models to outperform massive monolithic ones by focusing their "thinking" only on the most difficult parts of a task.
Why It Works: The Stability of the Gap
The reason these methods work is that most physical and logical processes have inherent "inertia." A robot moving toward a goal doesn't need to be told to "keep going" every millisecond if its path is clear.
By formalizing the "gap" between updates, researchers like Lukas Kesper and Sebastian Trimpe are even using Reinforcement Learning to learn the optimal triggering conditions Toward Multi-Agent Reinforcement Learning for Distributed Event-Triggered Control. Their work suggests that agents can maintain high performance while cutting communication by over 50% simply by learning which moments in a trajectory are "critical" and which are "routine."
Whether it's a ride-sharing system responding to a new request Chen and Cassandras or an LLM agent deciding when to interject in a Mafia game Eckhaus and Berger, the research direction is clear: the most intelligent systems are those that know when to stay silent.