← Back to Concept Contrast

Concept Contrast Fable

ReAct vs. planning: one reads the road while running, the other draws the map before setting out

Conclusion first:When building Agents, ReAct and planning are often compared in the same architecture discussion. ReAct lets the Agent alternate between reasoning and action, with the result of each step determining the next; planning lets the Agent first decompose the goal, lay out a sequence of steps, and then execute according to the plan. The difference between the two is not who is smarter, but tolerance for the unknown—ReAct reserves room for mid-course adjustment with incomplete information, while planning puts all the pressure of information completeness at the starting point.

When building Agents, ReAct and planning are often compared in the same architecture discussion. ReAct lets the Agent alternate between reasoning and action, with the result of each step determining the next; planning lets the Agent first decompose the goal, lay out a sequence of steps, and then execute according to the plan. The difference between the two is not who is smarter, but tolerance for the unknown—ReAct reserves room for mid-course adjustment with incomplete information, while planning puts all the pressure of information completeness at the starting point.

Fable

Two navigators' strategies for entering the mountains

A surveying team had to cross a pristine forest that had never been surveyed. The team had two navigators with completely different strategies for entering the mountains. The first navigator stopped every two hundred meters, climbed the tallest tree to observe the terrain ahead, and adjusted direction based on the ridge lines and stream directions he saw. His map grew more detailed the deeper they went. The second navigator spread out satellite images and contour maps at the mountain entrance and spent three days drawing a complete route from the entrance to the target point, marking seventeen turning points.

For the first three days, both arrived at the designated camps on time. The first navigator took some detours, but each detour was decided only after seeing a cliff ahead—the extra distance had a reason. The second navigator's route was straighter—at least straight on the map.

On the afternoon of the fourth day, the second navigator stopped at the eighth turning point. The shoal marked on the contour map was in reality a deep trench carved by the rainy season, wider than the bridging equipment could handle. He searched every map but could not find a viable alternative route, because all his decisions were based on the complete set of information from the mountain entrance, and that information had now dropped to zero.

At dusk the same day, the first navigator also reached this deep ravine. He retreated to the previous observation point, climbed the tree again to look at the surrounding mountain shapes, and chose a gentle slope on the east side to go around. It took two extra days of travel, but he never fell into the state of being unable to move once the plan broke.

After returning to camp, the survey team reviewed the boundaries of the two strategies. The second navigator's approach suited areas with complete information and a stable environment—road surveying, plains surveying—where information will not deceive you. The first navigator's approach suited areas with incomplete information and a changing environment—unfamiliar mountains, rainy-season jungles—where you must collect information at each step for the next.

The captain later assigned each survey team a "forward scout navigator" and a "route planner." The scout was responsible for opening a path in unfamiliar terrain using short-cycle observe-act loops, while the route planner was responsible for solidifying the scout's experience into repeatable standard routes once entering known terrain. The two did not replace each other, only relayed to each other.

The final lesson was printed on the title page of the surveying manual: if before setting out you have obtained all the true information about this land, draw the map first and then go; if this land has never been walked by anyone, look at the road in front of your legs at every step. Both maps can take you to the destination, but they assume different levels of map precision.

ReAct

An Agent reasoning pattern that alternates reasoning and action. The model first thinks, then executes a tool call, then thinks again based on the result—each step gathers information for the next. Suitable for dynamic tasks where environmental information is incomplete.

Planning

The Agent first generates a complete step-by-step plan before execution, then executes according to the plan. It depends on the completeness of the starting information and the stability of the environment. Suitable for structured, clearly bounded repetitive tasks.

Story Mapping

  • The first navigator climbs a tree to observe every two hundred meters before moving on: ReAct, an observe-reason-act loop at every step
  • The second navigator draws the full route at the mountain pass: Planning, generating a global plan before execution
  • A shallow shoal turning into a deep trench resets planning to zero: Planning lacks the ability to adjust midway when information changes abruptly
  • Outpost and planner relay: ReAct probes unfamiliar environments, Planning solidifies known paths

Translating into Agent behavior design

When designing Agent tasks, choose the pattern by scenario. For tasks with complete information and fixed steps (such as generating reports in a fixed format), use Planning to break down the steps in advance for maximum efficiency. For tasks with incomplete information or a dynamically changing environment (such as multi-tool search or open-ended research), use ReAct so that each step's result drives the next decision. For hybrid scenarios, use Planning to build the skeleton and ReAct to fill in the details: set the overall direction first, then let an observe-act loop adjust each segment internally.

京ICP备2026014272号