A Recipe vs. Improvising Off the Cuff

Opening Scene

Some dishes are cooked from a fixed recipe, every step decided in advance, executed start to finish without deviation. Others are improvised, the cook tasting and adjusting after each addition, deciding the next move only once the last one’s result is known. Both are legitimate ways to cook, suited to different situations. AI agents face this exact same choice, and it’s a foundational design decision, not an implementation detail.

In Plain English

Deliberative planning has an agent work out a full multi-step plan upfront, then execute it. Reactive, step-by-step approaches (often implemented via patterns like ReAct) have the agent decide only its next single action, observe the result, and then decide the action after that — repeating one step at a time rather than planning the whole sequence in advance. Each has genuine tradeoffs: upfront plans are more predictable and easier to review, while step-by-step approaches adapt more readily to information that only becomes available partway through the task.

The Old Way

Before this distinction was clearly recognized as a genuine, deliberate design choice, early agentic systems often didn’t make it explicit:

Early systems sometimes committed to a rigid, fully upfront plan even when new information partway through genuinely should have changed the approach.
Other early systems operated in a purely reactive, step-by-step manner with no upfront plan at all, sometimes wasting effort on approaches a small amount of upfront planning would have avoided.
There wasn’t yet a well-established vocabulary or framework for choosing deliberately between these two genuinely different approaches based on the task’s nature.

Recognizing planning style as a deliberate design axis — not simply an implementation detail — is a mark of real, accumulated maturity in agentic system design.

What’s Changing (and Why AI Is the Reason)

Well-designed agentic systems increasingly choose deliberately between upfront planning and step-by-step reactivity, based specifically on how predictable the task’s environment is.
Hybrid approaches have emerged that plan at a coarser, higher level upfront while adapting reactively within each individual planned step, connecting directly to the multi-step task decomposition covered in Article 5.
This connects directly to the reliability and evaluation concerns covered throughout this series — a fully upfront plan is generally easier to review and test before execution than a purely reactive one, which only reveals its behavior as it runs.

The Metaphor, Fully Extended

The Kitchen	Agent Planning Concept
Cooking from a fixed recipe, every step decided in advance	Deliberative planning: an agent’s full plan is worked out before execution
Tasting and adjusting after each addition	Reactive, step-by-step decision-making, adapting to each result
Choosing the right approach for the dish at hand	Choosing the right planning style for the task’s predictability
A recipe that’s easier to review before cooking even starts	An upfront plan that’s easier to review and test before execution

For Beginners: What to Actually Do

Practice identifying whether a given agentic task is predictable enough for upfront planning, or uncertain enough to require step-by-step reactivity.
Learn to trace through a reactive, step-by-step agent’s behavior one action at a time, noting how each result shapes the next decision.
Get comfortable reviewing an upfront plan for obvious flaws before letting an agent execute it, taking advantage of that approach’s inherent reviewability.

For Practitioners and Leaders: The Deeper Layer

Choose deliberately between deliberative planning and reactive step-by-step approaches based specifically on your task’s environmental predictability.
Consider hybrid approaches that plan coarsely upfront while adapting reactively within individual steps, for tasks that have both predictable and uncertain elements.
Recognize that upfront planning’s reviewability is a genuine reliability advantage, worth weighing against reactive approaches’ greater adaptability.

Quick Recap

Deliberative planning works out a full plan upfront; reactive approaches decide one step at a time based on each result.
Each has genuine tradeoffs: predictability and reviewability versus adaptability to new information.
Well-designed systems choose deliberately between these approaches based on the task’s environmental predictability.
Hybrid approaches combine coarse upfront planning with reactive execution within individual steps.

Where This Fits in the Series

Article 15 covered planning style as a deliberate design choice. Article 16 covers what happens when a single agent isn’t enough, and multiple agents need to coordinate within one system.