Search vs Planning

The Lens

Two fundamental strategies exist for navigating toward goals in complex systems:

Planning: Building a predictive model of the system, then deriving the optimal path before execution. Assumes you can shortcut to the answer without running every step.

Search: Executing actions, observing outcomes, updating direction. Accepts that you must run the system to know what happens.

The critical question: When does each strategy dominate?

This lens provides the framework for understanding when to think vs when to act, when to simulate vs when to touch reality, and why AI has shifted the balance toward search in many domains.

Computational Reducibility vs Irreducibility

The distinction between planning and search maps onto a deep computational concept:

Computationally Reducible Systems

Some systems have shortcuts. You can predict the outcome without simulating every step.

Planning works perfectly here: Think hard enough → derive correct answer → execute once.

Computationally Irreducible Systems

Some systems have no shortcut. The only way to know the outcome is to run the simulation step by step. The system is its own fastest simulator (Wolfram's concept).

Planning has a ceiling here: You can approximate, but there's always residual uncertainty that only running the system resolves.

The Key Insight

Most real-world domains of interest are computationally irreducible:

• Human behavior
• Markets
• Complex systems
• Relationships
• Creative work

We pretend they're reducible because iteration was expensive. Planning was the cope for high iteration costs. When you can only try once (surgery, rocket launch), you must plan extensively.

When Each Strategy Dominates

Planning Doesn't Fail in Irreducible Systems

Clarification: Planning doesn't fail—it produces useful approximation. You can still reason about the system, predict rough behavior, identify likely regions.

What planning can't do in irreducible systems: Give you the exact answer. There's always a gap between your model and reality.

The relationship:

• Planning gets you close (reduces million possibilities to hundred)
• Search closes the gap (finds the one that actually works from that hundred)

Why Planning + Search Together

Neither strategy works alone:

Planning Alone

Assumes perfect model → stuck at wrong answer in irreducible systems

If your model is wrong (and in irreducible systems, it always is to some degree), you converge on the wrong answer with high confidence.

Failure mode: "But my analysis said this would work!" — It did, based on your model. Reality runs on different rules.

Search Alone

Random walk → intractable in high-dimensional spaces

Pure search with no guidance is exponentially slow. In a space with 1000 dimensions, random search takes forever to find anything.

Failure mode: Trying everything randomly, never converging because no direction.

Together: Initialization + Descent

Planning = good initialization (where to start, which direction is likely "warmer") Search = descent (iterate toward optimum using reality feedback)

This is gradient descent:

• Bad initialization + good descent = slow convergence
• Good initialization + no descent = stuck at starting point
• Good initialization + descent = fast convergence

Planning constrains search space. It doesn't need to be correct—just constraining. Narrowing from 1 million options to 100 makes search tractable.

AI's Effect: Collapsing Iteration Cost

AI fundamentally shifts the balance by collapsing iteration cost.

Before AI

High iteration cost → planning dominates → "think before you act"

After AI

Low iteration cost → search dominates → "act to learn"

The Mechanism

AI doesn't make systems more reducible. It makes running them cheaper.

So you stop pretending you can predict, and start actually running experiments.

$\text{Strategy} = f(\text{iteration cost}, \text{system reducibility})$

AI reduced iteration cost → search became viable → search now dominates in more domains.

Reality Contact as the Differentiator

Planning = simulated reality contact (running the model in your head) Search = actual reality contact (running the system in reality)

AI Accelerates Reality Contact

AI makes interfacing with reality faster:

• Process feedback quickly (what did that error mean?)
• Generate next iteration quickly (try this variation)
• Extract gradient from binary outcomes (see gradient extraction)

This makes search the dominant strategy:

Before AI: Reality contact slow/expensive → planning cheaper → plan first With AI: Reality contact fast/cheap → search more efficient → act first

The Danger

AI without reality contact = faster simulation = amplified blind spots

You can plan more elaborately, simulate more scenarios, build more sophisticated models—all in pure simulation. AI accelerates this too.

AI with reality contact = faster search = accelerated truth convergence

Each iteration touches reality, gets feedback, updates model. AI accelerates this loop.

The tool is neutral. The question is: Are you using AI to avoid reality or interface with it faster?

The Planning-Search Loop

The optimal strategy in irreducible domains is a loop, not a choice:

1. PLAN (LLM constrains search space)
   → Hypothesis about what might work
   → Direction to search

2. ACT (execute in reality)
   → Try the hypothesis
   → Touch actual territory

3. OBSERVE (collect signal)
   → What happened?
   → How does it differ from prediction?

4. EXTRACT GRADIENT (LLM converts outcome to direction)
   → Binary outcome → directional information
   → What does this suggest about where to go next?

5. UPDATE PLAN (refine model)
   → Incorporate new information
   → Narrow search space further

6. REPEAT until converged

This Loop Is Irreducible

You cannot skip steps:

This is the minimal structure for learning in irreducible domains.

Variations Are Parameter Changes

Variations on the loop exist, but they're parameter changes, not structural alternatives:

Agentic AI as Planning-Search Loops

Agents aren't "LLMs that call tools." Agents are planning-search loops with reality contact.

What Makes an Agent

Infrastructure Determines Viability

Loop iteration cost determines if search is viable.

Without infrastructure: Each iteration expensive (manual wiring, state management) → planning dominates With infrastructure: Iterations cheap (automated loops, persistent state) → search dominates

This is what good AI infrastructure enables: Cheap iteration on planning-search loops so search can dominate.

High-Leverage Domains for Search

Search dominates in domains with:

1. Symbolic solution space - Structured rules, defined moves
2. Verifiable outcomes - Clear right/wrong, testable
3. Complex search space - Too many options for exhaustive planning

Common pattern: Binary domains where you can insert gradient extraction to create optimization loops.

Practical Application

When to Plan More

• High iteration cost (can only try once)
• Catastrophic failure (mistakes are irreversible)
• Simple model (you can actually predict)
• No feedback available (can't observe outcomes)

When to Search More

• Low iteration cost (can try many times)
• Recoverable failure (mistakes are learning)
• Complex model (too many unknowns to predict)
• Fast feedback (know quickly if it worked)

The Meta-Heuristic

If you're spending more than 20% of time planning without reality contact, you're probably in the wrong strategy.

AI reduced iteration cost. Adjust your planning/search ratio accordingly.

Related Concepts

• AI as Accelerator - AI collapses iteration cost, shifts domains toward search
• Reality Contact - Search requires touching territory, planning stays in simulation
• Gradients - Gradient extraction converts binary search outcomes to directional signal
• Algorithmic Complexity - Planning = reducing complexity through modeling
• Cybernetics - The planning-search loop is a cybernetic feedback system
• Startup as a Bug - Startups as search in computationally irreducible market space
• Clarity - Planning produces initial clarity, search refines it
• Skill Acquisition - Skill building is search through technique space with reality feedback

Key Principle

Planning assumes computational reducibility. Search accepts irreducibility. Most real domains are irreducible—plan to constrain, search to converge.

Planning produces useful approximation but has a ceiling in irreducible systems. Search closes the gap between approximation and truth through reality contact.

Neither alone works: Planning alone → stuck at wrong answer with confidence. Search alone → random walk, never converges.

Together: Planning provides initialization (constrains search space), search provides descent (finds actual optimum).

AI collapses iteration cost → shifts more domains from planning-dominant to search-dominant. The strategy that seemed to require extensive planning (code, writing, design) now benefits from rapid iteration.

The planning-search loop (plan → act → observe → extract gradient → update → repeat) is the irreducible algorithm for navigating irreducible systems. You cannot skip steps. Variations (parallel, hierarchical, batch) are parameter changes, not structural alternatives.

Agentic AI = planning-search loops with reality contact and persistent state.

Use AI to accelerate the loop, not to avoid it.