Reactive Agent

A Reactive Agent selects actions based solely on its current percepts and a fixed set of condition-action rules, without maintaining internal state or reasoning about the future. The classic example is a thermostat (if temperature < setpoint, turn on heater) or a vacuum-cleaner agent with bump sensors (if obstacle, turn). Reactive agents are computationally cheap (<1 ms per decision), provably correct under bounded conditions, and cannot suffer from stale world models — but they cannot handle tasks requiring memory of past events or long-term planning. In AI practice, reactive designs are used for low-level control loops (motor PID controllers, network load balancers) while

A reactive agent responds immediately to inputs without planning — stimulus → response. It's fast (<100ms) but shallow. Use for simple, repetitive tasks where latency matters more than sophistication: API call routing, simple Q&A, first-pass content moderation. LLMs in reactive mode skip deliberation.

Reactive agent deployment: (1) for latency-critical tasks (<500ms SLO), use small models (1.5B–3B) in reactive mode (no CoT, no tool use), (2) for complex tasks, use reactive as first-pass filter — if confidence < threshold, escalate to deliberative agent with larger model, (3) reactive agent prompt: minimal system prompt, clear task, examples in context, (4) monitor reactive_handled_ratio — if >90% of requests are handled reactively, your system is efficient.