Research Deep-Dive

Research — LLM Agents.md

신뢰도 높음원본 요약편집: Cairni · 방금 · AI 생성v1

Source Overview

A set of personal research notes exploring the practical mechanics of LLM agents — what patterns work beyond a single prompt, and where they tend to break. Research — LLM Agents.md


Central Question

*What actually makes an LLM "agent" work in practice, beyond a single prompt?*

The notes are organised around four sub-topics: the ReAct pattern, tool use, memory, and the planner-vs-reactive debate.


Sections at a Glance

SectionCore claimNotable flag
ReActInterleaves reasoning with tool calls; grounding beats pure chain-of-thoughtCan loop / repeat failing actions
Tool useFunctions extend the model beyond training; schema structure reduces errorsToo many tools hurt selection accuracy
MemoryShort-term = context window; long-term = retrieval (RAG)Contradiction flagged (see below)
Planning vs reactivePlanners = predictable but brittle; reactive = adaptive but may wanderOpen question on when to prefer each

Flagged Contradiction

The notes explicitly mark a disagreement on memory approaches:

  • One note claims most "agent memory" is just retrieval + summarisation, and that naive vector search misses temporal and structural relationships.
  • A cited paper claims embeddings alone suffice.

These positions conflict and are unresolved. Research — LLM Agents.md


Key Entities & Concepts Mentioned

  • ReAct — reason-then-act loop pattern
  • Chain-of-thought — baseline reasoning approach ReAct improves upon
  • Tool use / function calling — search, code execution, external APIs
  • Structured/forced tool schemas — proposed fix for tool-selection errors
  • RAG (Retrieval-Augmented Generation) — primary mechanism for long-term memory
  • Vector search / embeddings — memory retrieval mechanism (disputed sufficiency)
  • Explicit planners — decompose-then-execute agent architecture
  • Reactive loops — ReAct-style adaptive execution

What This Source Contributes

This source provides a framing scaffold for the research wiki: it names the four main sub-topics, surfaces a concrete contradiction worth investigating (embeddings vs richer memory), and poses the open planning question as an explicit research gap. It does not itself resolve any of these questions — resolution (if any) must come from the underlying articles and papers referenced in the notes.