Agent Patterns Hub

Adaptive Search Strategies for AI Agents: Beyond Static Queries

Most agent search strategies are static: define a search plan, execute it, return results. This works when you know what you are looking for and approximately where it is. It fails when the search space is large, the target is vaguely defined, or the best search strategy cannot be determined in advance.

Adaptive search replaces fixed plans with dynamic strategies that get smarter as they run.

The Satisficing Principle

Herbert Simon's satisficing theory: instead of searching for the BEST answer (which grows exponentially expensive), search for the first answer that exceeds a "good enough" threshold. The threshold itself adapts based on what the search finds.

When finding great results, raise the threshold (demand better). When finding nothing, lower the threshold (accept what is available). The threshold never drops below 0.3 (garbage) or exceeds 0.95 (perfectionism trap).

Three Types of Information Scent

  1. Surface scent (keyword density): How many relevant terms appear in this area? Cheap but shallow. A glossary has high keyword density but low information content.
  2. Deep scent (semantic proximity): How closely does the content match the question's meaning? Requires embedding similarity or LLM judgment. Catches content using different vocabulary.
  3. Structural scent (cross-reference frequency): How many other parts reference this area? A file imported by 50 others is almost certainly more important than one with matching keywords but zero references.

Combine them: composite = 0.3 * surface + 0.5 * deep + 0.2 * structural. Allocate more agents to high-scent areas, abandon low-scent areas.

The Diminishing Returns Gate

Every round of adaptive search should check: what fraction of findings this round are genuinely new? If new_info_ratio drops below 10%, stop. More rounds will not help.

This prevents the common failure mode of search systems: running forever because there is always one more place to look.

When to Use Adaptive vs Static Search

ScenarioBest Approach
Target is in the obvious placeStatic (both work equally)
Target location is unknownAdaptive (scent guides toward it)
Multiple valid answers existAdaptive (raises threshold, finds better)
No good answers may existAdaptive (lowers threshold, returns best)
Large space, tight budgetAdaptive (concentrates budget on high-scent areas)

Adaptive Search is pattern 9 of 15 in the Protocol Playbook.