Agent vs RAG vs Pipeline
These three terms are conflated everywhere. The distinction:
- →Pipeline — fixed sequence of steps. Predictable, fastest, cheapest.
- →RAG — pipeline with one retrieve step. Grounds answers in your data.
- →Agent — the LLM decides the steps at runtime, including calling tools and looping.
When to use which
If the task is the same every time, build a pipeline. If you need fresh facts, add RAG. If the task varies and needs decisions or tool use, you actually need an agent.
The agent loop
Every agent, no matter the framework, runs some variant of this loop:
- →Receive goal / next user message
- →Think — LLM reasons about what to do next
- →Act — call a tool (function call) or produce a final answer
- →Observe — read the tool's output
- →Repeat from 'Think' until done
Core primitives
Modern agents are built from a small set of repeating ideas:
- →Function calling — JSON-schema-based tool invocation; the bedrock primitive
- →Memory — short-term (conversation), long-term (vector store), MemGPT-style managed
- →Reasoning patterns — ReAct, Plan-and-Execute, Tree of Thought, Reflexion
- →Guardrails — permission gates, budget limits, human-in-the-loop approvals
- →Observability — trace every LLM call, tool call, retry
Which framework should you use in 2026?
The agent framework landscape has settled around four tools:
- →LangGraph — state machine + graph for agents; the modern LangChain answer
- →CrewAI — role-based crews; great for hierarchical multi-agent teams
- →OpenAI Agents SDK — handoffs, guardrails, built-in tracing; OpenAI's official stack
- →MCP (Model Context Protocol) — Anthropic's open standard for tool servers