Open Questions

This page collects what remains unresolved or actively debated across the notes gathered on LLM Agents. Each item points back to the sub-topic where the tension arises.

• When should you use a planner vs a reactive loop? The notes identify two camps — explicit planners (decompose-then-execute) and reactive loops (ReAct-style) — but offer no settled answer for when each is the right choice. Planners are more predictable yet brittle when the environment changes mid-task; reactive loops adapt but can wander without converging. [Research — LLM Agents.md]

• Do embeddings alone suffice for agent memory, or is retrieval + summarization required? One note claims that most practical "agent memory" is just retrieval + summarization via RAG, and that naive vector search misses temporal and structural relationships. A conflicting paper argues embeddings alone are sufficient. This disagreement is flagged as an active contradiction and has not been resolved. See Agent Memory for the full conflict. [Research — LLM Agents.md]

• How do you prevent ReAct loops from getting stuck? The ReAct pattern is noted to sometimes loop or repeat a failing action without escaping. No mitigation strategy is documented in the current notes. [Research — LLM Agents.md]

• At what point does adding more tools degrade selection accuracy? One source argues that structured/forced tool schemas cut errors sharply, while another warns that too many tools degrade the model's ability to select the right one. The threshold or tradeoff curve is not yet characterized. See Tool Use & Function Calling. [Research — LLM Agents.md]

• How can agent memory better capture temporal and structural relationships? Current retrieval approaches (vector search) are noted to miss these relationship types. No concrete alternative approach is documented in the notes. [Research — LLM Agents.md]

• LLM Agents — Research Overview
• ReAct Pattern
• Tool Use & Function Calling
• Agent Memory
• Planner vs Reactive Agent Architectures
• Retrieval-Augmented Generation (RAG)
• Chain-of-Thought Reasoning
• Research — LLM Agents.md (source)