Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-25 · high
JoeJag's post separates model, inference service, harness and agent system, and explains why identical Sonnet weights act differently across tools.
Why now: The split matters as more builders stack MCP servers and skills onto the same handful of shared models without a common vocabulary for what breaks where.
The word 'Claude' hides four separate things, per JoeJag's breakdown of the agent stack. The model is just the trained weights. Sitting on top of that, the inference service is the hosted runner that meters API calls. The harness is the interaction layer where MCP servers, skills and tool routing live, and the agent system is all three stacked together.
The load-bearing claim is that MCP and skills are harness features, not model capabilities. That's why the same Sonnet weights behave differently in Claude CLI, in Cursor, and in a plain LangChain script. Each tool wires up its own tool calls and skill routing on top of the same weights.
The layers matter once something breaks. JoeJag frames it as deciding whether a failure belongs to your prompt, your tool wiring, or the model itself. Each answer points at a different fix.
A tool-routing bug in a harness won't go away no matter how the prompt gets rewritten. A real model limitation won't get fixed by swapping MCP servers either.
Each link below shares sources, entities, or timing with this story.
The study covered GPT-4o, Claude 3.5 Sonnet and Llama-3.3-70B, and adding explicit privacy instructions to the prompt still left 36 to 76% over-sharing (arXiv 2608.24957). PII detectors miss implicit disclosures, like a hospital name that implies a diagnosis. The middleware in...
A GitHub repo cataloging Claude Code tips doesn't normally warrant a top story. But shanraisshan/claude-code-best-practice at 53.4K stars isn't a tips list anymore. It's the de facto reference for how an entire generation of developers is learning to work with AI coding agents...
The defense-first MCP pattern from Christian Schneider treats every tool schema like a request from a stranger. A control point outside the client runs a five-stage validation pipeline. Stages 01 through 04 gate the discovery path, inspecting every schema before the model ever...
Figma released its MCP server in open beta, and this one matters more than most MCP launches. AI agents can now read, create, and edit components directly on the Figma canvas using your design system variables. The use_figma tool works with Claude Code, Cursor, Codex, Copilot...
Raj Nagulapalle's FetchSandbox MCP took 107 votes on August 23, wiring 70+ API sandboxes into Cursor or Claude Code via MCP config. The claim is narrower and more testable than most agent tooling: reproduce the real integration failure against a sandbox, apply the fix, re-run...
PostHog Desktop launched August 26 running a fleet of coding agents with Claude and GPT models, plan mode, parallel execution, MCP servers and a skill marketplace, whose context is the product's own production signals: in-app activity, logs, errors, payments and session record...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.