The Great Unbundling of RAG: Why AI Agents Are Building Their Own Context

RAG Pipeline vs Agent Context Construction - linear retrieval pipeline compared to an agent with multiple tools constructing context iteratively RAG is a delivery service. Agents are researchers. The difference matters.

We spent two years perfecting our RAG pipeline - chunking, embedding, reranking. It felt like the pinnacle of context delivery. Then, bit by bit, the ground started shifting. Cursor’s Planner, Devin’s workspace memory, Copilot Workspaces - all hinting at a new pattern. And when Anthropic launched Skills, it finally clicked: our AI began reading markdown files from a filesystem instead of querying our beautifully tuned vector store.

At first, it felt like someone bypassing your front door to climb through a window. Then I realized - this isn’t a bug. It’s the future.

The Problem We’ve Been Ignoring

RAG became the default answer to “how do I give my AI context?” But RAG is just a delivery service. You pre-chunk, embed, retrieve 5 snippets, and hope the model can reason with them. Human experts don’t work that way. When I’m hunting threats, I don’t want the 5 most semantically similar logs. I want to:

Grep for specific IOCs
Run SQL queries
Read procedural docs
Follow trails iteratively

I need tools, not chunks.

What’s Actually Emerging

This new generation of agents - from Cursor to Claude - is converging on the same idea: give agents a workspace, not just retrieved text.

A Skill is just a folder with markdown files and scripts. The AI sees a table of contents (≈50 tokens), opens what it needs, runs code when helpful.

Progressive disclosure instead of pre-chunking. Exploration instead of retrieval. And because models are trained on codebases, they’re already good at this.

The Architecture That’s Taking Shape

Context Source Architecture - four quadrants showing Filesystem, Vector Search, SQL, and Agent Orchestration as complementary context sources Vector databases aren’t dying. Their monopoly on context is.

It’s not filesystems vs vector stores - it’s context construction vs context retrieval:

Filesystem → procedural knowledge, workflows
Vector search → semantic similarity
SQL/structured → precise filtering
Agent orchestration → navigate all three

Vector databases aren’t dying. Their monopoly on context is.

The Real Shift

The moat is moving from how you store context to how you architect agent autonomy.

What tools do you expose?
How do you structure explorable knowledge?
What guardrails prevent hallucinated queries?
How do you make agent trails auditable?

We perfected RAG in 2024. 2025 is about unbundling it.

Are your agents still asking for chunks - or building their own context?