The ChatGPT routing strategy (GPT-5-main vs GPT-5-thinking) mirrors what I learned the hard way with my automation agent. Started using Opus for everything. Burned through API limits. Switched to Haiku by default. Quality actually improved because simpler tasks got clearer instructions instead of over-engineered prompts. Now I route based on task complexity: Haiku for email/scraping, Sonnet for content, Opus when multi-step reasoning required. The classifier layer matters more than the model itself. Cost dropped 70%, no quality loss. https://thoughts.jock.pl/p/claude-model-optimization-opus-haiku-ai-agent-costs-2026
This is a super clean mental model: MCP = tool interface (plumbing), RAG = knowledge injection (memory), Agents = decision loop (manager). They’re not rivals, they sit at different layers of the stack.
The next real battleground is trust + security: once MCP connects models to real systems (files, Slack, GitHub, internal APIs), the cool demo becomes an attack surface, so teams will need least-privilege tool access, audit logs, and safety checks between agents and external MCP servers
What about this use case, how is the end result different? I have MCP connected to Notion which stores files I want to reference or I have those as md files saved in my directory in Claude Code. Both seem like different paths to RAG
Yeah, both can feed RAG and the final answer can look basically the same. The difference is the control plane: local md is inside your boundary so you control chunking, indexing, caching, and what even counts as searchable. Notion via MCP is live tool access with auth, rate limits, permission bleed risk, and you really need audit logs when something goes wrong. Long term the plumbing gets standardized, but the real wins and failures still come from your retrieval rules plus safety checks, not the connector.
Great breakdown, especially the clarification that MCP, RAG, and Agents operate at different layers rather than competing. The routing explanation was also super helpful. Clear, practical refresher on how modern AI systems are actually structured.
These are all RAG. Every one of them. Embedding search and retrieval doesn’t define RAG, it’s just one example of it.
MCP use exposes you to EchoLeak style compromises with no ability to even know it happened. If you don’t have a dedicated MCP security agent between your agents and third party MCP’s you might as well just resign now.
A practically useful comparison would be between MCP, Legacy Tool Calling, and Semantic Routing.
The ChatGPT routing strategy (GPT-5-main vs GPT-5-thinking) mirrors what I learned the hard way with my automation agent. Started using Opus for everything. Burned through API limits. Switched to Haiku by default. Quality actually improved because simpler tasks got clearer instructions instead of over-engineered prompts. Now I route based on task complexity: Haiku for email/scraping, Sonnet for content, Opus when multi-step reasoning required. The classifier layer matters more than the model itself. Cost dropped 70%, no quality loss. https://thoughts.jock.pl/p/claude-model-optimization-opus-haiku-ai-agent-costs-2026
This is a super clean mental model: MCP = tool interface (plumbing), RAG = knowledge injection (memory), Agents = decision loop (manager). They’re not rivals, they sit at different layers of the stack.
The next real battleground is trust + security: once MCP connects models to real systems (files, Slack, GitHub, internal APIs), the cool demo becomes an attack surface, so teams will need least-privilege tool access, audit logs, and safety checks between agents and external MCP servers
What about this use case, how is the end result different? I have MCP connected to Notion which stores files I want to reference or I have those as md files saved in my directory in Claude Code. Both seem like different paths to RAG
Yeah, both can feed RAG and the final answer can look basically the same. The difference is the control plane: local md is inside your boundary so you control chunking, indexing, caching, and what even counts as searchable. Notion via MCP is live tool access with auth, rate limits, permission bleed risk, and you really need audit logs when something goes wrong. Long term the plumbing gets standardized, but the real wins and failures still come from your retrieval rules plus safety checks, not the connector.
Got it. Thanks!
I like the point of view. I am actually more interested in consequences about art.
Great Content !
Great breakdown, especially the clarification that MCP, RAG, and Agents operate at different layers rather than competing. The routing explanation was also super helpful. Clear, practical refresher on how modern AI systems are actually structured.
Maybe I recall it wrong: I have seen “12 Architectural Concepts Developers Should Know” in previous newsletters.
Really great content. Good explanation about Agents vs MCP vs RAG.
If only it were that tidy.
These are all RAG. Every one of them. Embedding search and retrieval doesn’t define RAG, it’s just one example of it.
MCP use exposes you to EchoLeak style compromises with no ability to even know it happened. If you don’t have a dedicated MCP security agent between your agents and third party MCP’s you might as well just resign now.
A practically useful comparison would be between MCP, Legacy Tool Calling, and Semantic Routing.
This is “I want to pretend ai know”