Agent planning beyond a ReAct loop

WHAT THIS TESTS This probes whether you understand the spectrum of agent control strategies and can articulate the core tension: adaptability versus efficiency and predictability.

A GOOD ANSWER COVERS ReAct interleaves reasoning and acting: the agent thinks, takes one action, observes the result, and repeats. This is highly adaptive because each step responds to fresh information, but it spends one or more LLM calls per step, so a long task becomes slow and expensive, and the agent can drift or loop without a global plan. Plan-then-execute first asks the model to generate a complete plan of steps, then executes them, optionally with a cheaper executor. This cuts expensive reasoning calls, gives an inspectable plan, and improves coherence on complex tasks, but it is brittle: if reality diverges from the plan, an early failure cascades unless you add re-planning when a step fails. Hierarchical decomposition sits between: a high-level planner breaks the goal into subgoals, and sub-agents or sub-plans handle each, possibly recursively. This manages complexity and keeps each level focused, balancing structure with the ability to re-plan a failed subtask without discarding everything. The trade-off summary: reactive maximizes adaptability at high cost; upfront planning maximizes efficiency and coherence but risks brittleness; hierarchical approaches seek a middle ground with added orchestration complexity.

COMMON WRONG ANSWERS Claiming upfront planning is always better; ignoring that plans break when the environment changes; treating ReAct as free; conflating the three strategies; forgetting that re-planning is what rescues plan-then-execute.

LIKELY FOLLOW-UPS How do you trigger re-planning? When is ReAct's adaptability worth its cost? How does hierarchy reduce context length per call? How do you bound total cost?

ONE CONCRETE EXAMPLE For a research-and-report task, a pure ReAct loop might make dozens of expensive calls and wander. Instead a planner decomposes it into gather sources, extract findings, and write report subgoals; a sub-agent handles each, and if the source-gathering subgoal returns nothing useful, only that subtask is re-planned, saving cost while staying adaptive.