Planning Depth vs. Reactivity: How Agents Decide When to Think

Agent architectures sit on a spectrum from fully deliberative to fully reactive. Deliberative planners generate a complete task plan before taking any action — useful when the environment is predictable and actions are expensive or irreversible. Reactive agents act immediately on the latest observation with no lookahead — useful in fast-changing environments where stale plans are harmful. Most production agents use a hybrid: a high-level plan generated upfront with reactive replanning triggered by unexpected tool results. The implementation mechanisms matter. Task decomposition quality is the dominant factor in plan reliability: poorly decomposed tasks lead to agents that complete subtasks technically but miss the goal. Hierarchical planning — breaking a goal into high-level phases, then decomposing each phase into concrete steps — produces more robust agents than flat task lists. Plan verification is underused: asking the model to critique its own plan before executing it catches a significant fraction of logical errors and missing steps at near-zero cost. Interruption policies are critical for long-running agents: defining in advance which types of unexpected results should trigger a full replan (vs. a local retry) prevents agents from either thrashing on every minor obstacle or blindly proceeding through cascading failures. Benchmarks like SWE-bench (software engineering tasks) and WebArena (web navigation) show that planning quality, not model size alone, is the primary driver of autonomous task completion rates.