← Back to concept explainers

Concept Fable

Planning

Conclusion first:Planning is the process by which an AI system breaks goals into steps, chooses actions, and adjusts its route as conditions change. It is closer to real delivery than single-pass reasoning: it needs to handle goals, constraints, resources, dependencies, risks, approvals, and failure branches. The planning quality of an enterprise Agent determines whether subsequent execution is controllable, especially when tool calls, human approvals, budget limits, and external system states are constantly changing. Good planning is not a one-time fixed route, but a judgment framework that leaves room for adjustment during execution.

Planning · LLM Agents

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The salt merchant did not raise the sails immediately before going to sea

The salt merchants of Cape Town used to run only one coastal route. The captain watched the wind, counted the crew, loaded the salt sacks, raised sail at dawn, and could return by evening. The old people said that at sea the most important thing was quick hands; thinking too much would make you miss the tide.

Later, the salt merchant received a large order that required passing through three hidden reefs, two tax ports, and a stretch of frequently foggy strait. The buyer demanded punctuality, the cargo hold could not get damp, and the accountant also stipulated that fuel and labor budgets must not be exceeded.

The young captain still wanted to set off as usual. When the ship reached the first reef, he discovered the tide level was wrong; after detouring, he missed the tax port's opening time; to make up time, he hired two more sailors, and the bookkeeper returned the budget sheet that same day.

The chamber of commerce gave a patch: every boat going to sea had to memorize a fixed chant. First look at wind, then water, then cargo, then port. The chant was not wrong, but when fog locked the strait, a tax port temporarily closed its gates, or a buyer urged early delivery, it did not tell the captain which route to abandon or which promise to keep.

The old navigator called the captain to the sand table. He did not sail the ship for the captain, but first asked five things: what the final deliverable is, which conditions cannot be broken, which resources are limited, which risks would change the route, and which actions must wait for approval.

They split the voyage into segments and wrote each segment's goals, optional routes, trigger conditions, and costs on wax tablets. Whether to wait or detour when the fog is thick, which port to switch to when a tax port closes, and whether to protect the delivery date or the cost first when the budget is exceeded—all the bases for judgment were written down in advance.

The second voyage still ran into trouble. The fog was thicker than forecast, a sailor fell ill, and the buyer changed the unloading window again. But the captain didn't panic. What he saw was not a pile of trouble, but a set of steps that could be reordered, replaced, and traded off.

The salt ship ultimately arrived half an hour late, but it preserved the cargo, the budget, and the buyer's critical window. When the chamber of commerce reviewed it, they found that victory did not depend on whether any single route was perfect, but on having thought through the goal, the constraints, and the ways things might change before setting out. From then on, Cape Town no longer treated preparation as delay. On a complex voyage, breaking down the goal, choosing the route, and factoring change into your judgment is itself part of the work.

Reveal

This story is about: planning

Planning is the process by which an AI system breaks goals into steps, chooses actions, and adjusts its route as conditions change. It is closer to real delivery than single-pass reasoning: it needs to handle goals, constraints, resources, dependencies, risks, approvals, and failure branches. The planning quality of an enterprise Agent determines whether subsequent execution is controllable, especially when tool calls, human approvals, budget limits, and external system states are constantly changing. Good planning is not a one-time fixed route, but a judgment framework that leaves room for adjustment during execution.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Deep-sea salt delivery: tasks with clear enterprise goals but complex paths
  • Fixed formulas: surface-level methods that only give the model templates or fixed steps
  • Reefs, tax harbors, fog seas, budgets: constraints, risks, external dependencies, and cost boundaries during execution
  • Five questions before the sandbox: planning needs to clarify goals, hard constraints, resources, risks and approval points
  • Optional paths on the wax tablet: step decomposition, conditional branching, and action selection
  • What changes on the second voyage: the plan must be able to adjust to new conditions
  • Protect the critical window: planning serves business outcomes, not generating the most beautiful route
  • Reviewing the Voyage: Planning Outputs Need to Be Inspectable and Improvable

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

agenttool usetask decompositionreasoningworkflow
京ICP备2026014272号