From RAG to context engineering: how agent retrieval evolved through three eras
mdancho84 · x · 2026-10-13
Matt Dancho's thread argues context engineering isn't just RAG rebranded — both solve the same problem of building the right context for your LLM, but the paradigm has shifted through three stages:
- RAG (2020–2023): fixed one-shot pipeline — retrieve, inject, generate. Rigid; the LLM had no say in its context window.
- Agentic RAG (2023–2024): retrieval becomes a tool the agent can call or skip, with multi-hop retrieval and index routing. The model starts deciding whether and what to retrieve.
- Agentic search in context engineering (2025+): the agent combines file search, database queries, semantic search, web search, memory retrieval, and shell tools to dynamically assemble the context window per query.
The thread includes architecture diagrams comparing all three stages. Why data scientists should care: it changes how you pick data sources, design retrieval tools, manage token budgets, and debug bad retrievals. RAG isn't dead — it's now just one tool in a much larger system.
More from coding & agent
- Stack Overflow Expands Stack Internal to Give AI Agents 'Trusted' Enterprise Knowledge — pchandrasekar · 2026-10-13
- Grok Bot's new X access, email and Opus turn personas into always-on agents, says creator Peter Yang — petergyang · 2026-10-13
- jscpd adds semantic cross-language clone detection to map AI code migrations — Affectionate-Blood92 · 2026-10-13
- Benchmarking Microsoft's MXC sandbox: git jumps from 10ms to 1.24s without a writable temp dir — lemire · 2026-10-13
- Let your coding agent auto-clean git worktrees and branches after deploy — NielsRogge · 2026-10-13
- Generative art built with Opus 5.5 writing Three.js code — techartist_ · 2026-10-13