Search, Heuristics, and Best-First Planning for LLM Agents in Agentic AI
Search, Heuristics, and Best-First Planning for LLM Agents
Planning as search
Many planning problems are search problems: you explore possible sequences of actions and pick the best.
Heuristics you can implement today
- Constraint satisfaction first
- Minimize tool calls
- Prefer high-confidence sources
- Stop early when good enough
Best-first planning (intuitive)
Generate candidates, score them, expand the most promising one next. It’s like ToT but applied to action sequences.
Hybrid scoring
Score = rules (hard checks) + LLM judge (soft ranking). This mix is how production systems stay both safe and flexible.
Avoiding over-search
Set depth, branching, and budget. Without limits, the agent becomes a search engine that never returns an answer.
