Six to eight months, zero engineers, 24 shipped AI skills. If you've ever wondered whether "vibe coding" scales past side projects, today's lead story is the closest thing to a field report you'll get.

## The lead story: planners, not programmers, built the agents

Matt Ragone leads AI adoption at Range, a wealth management firm. On the AI Engineer podcast, he laid out what happened when his team of roughly 20 planners (not developers, planners) started building AI skills on top of real client data. Six to eight months in: 24 skills across the company, 17 of them for financial planning, built by planners, marketers and client-success staff who never wrote a line of code.

The unlock wasn't a better prompt. Ragone says two things had to be true first: the agent needed access to actual client data (accounts, insurance docs, mortgages) inside Range's platform, and it needed to know how Range delivers advice, not generic advice. The second one turned out to already be solved by accident — Range's planners had standardized their recommendation process in Notion pages and Google Docs long before any agent existed. When the tooling arrived, the team just pointed Claude at that existing documentation: "instead of reinventing the wheel, it was 'Hey Claude, just reference this document as part of your review process.'"

That's the unglamorous finding worth stealing: the documentation habit, not the model, is what made this fast.

Ragone's five-step framework for anyone building a skill: pick a task you do daily that's time-consuming but simple to automate (not the flashiest one), teach it with 2-5 real cases including the edge cases you had to correct, pack it with clarifying questions and hard limits, test it on 5-10 real cases before shipping, ship it with a named owner and a separate reviewer. His warning: "Your first instinct is going to be to pick the flashiest thing, the most complex thing. I'm telling you from experience, do not do that."

The compliance angle is a design constraint, not a footnote — every skill at Range is blocked from sending anything to a client until a CFP reviews it. The agents draft; humans check.

And here's the part that should make every team racing to ship agents pay attention: the original bottleneck (how fast can a planner type during a call) is basically solved. The new bottleneck is maintenance. "The bottleneck has completely moved," Ragone said. "Now we're starting to ask, who is making sure all these get tested, updated?" His fix right now is one owner and one reviewer per skill — which works fine at 24 skills and may not survive at 100.

[YOUR TAKE: does your "skill library" have an owner, or did you build it and walk away?]

## Quick hits

**A four-key pad just for AI coding.** Keychron's new F4 HE 8K Vibe Coding Keypad is a $35 macro pad with four programmable keys aimed at Codex and Claude workflows — send prompt, voice input, accept/reject response. Magnetic switches, 8000Hz polling, preset Claude/Codex layouts in the box. [KitGuru](https://www.kitguru.net/peripherals/keyboards/joao-silva/keychron-launches-four-key-keypad-for-ai-assistants/)

**Endpoint security now watches your agents and MCP servers.** Palo Alto shipped Cortex XDR One on October 19, treating AI agents, MCP servers and browser extensions as things an endpoint tool must see. Ten days later Microsoft made local AI-agent and MCP-server discovery generally available in Defender, covering more than 50 agents on Windows and macOS. No independent lab has tested either yet, and neither vendor publishes public pricing for these specific features. [tech-insider.org](https://tech-insider.org/mcp-security-cortex-xdr-one-vs-microsoft-defender-ai-agents-2026/)

**A faster way to run Jev-style decision models.** Flide, an open-source inference engine, claims up to 9.5x lower latency serving structured decision models versus vLLM, SGLang and qev — though its own benchmark table shows bigger swings for specific models (28.1x for one model on one workload, near 1.0x for another). These are the project's own numbers, run on a single H200, with no independent rerun yet. Worth a look if you're experimenting with structured-output decision models in an agent pipeline. [news.lavx.hu](https://news.lavx.hu/article/flide-claims-up-to-9-5x-lower-latency-for-jev-style-decision-models)

## What this means for you

- Before you build an agent for a repeatable task, write down how you already do it well. Range's whole rollout ran on docs that existed before the AI did.
- Pick the boring, repetitive task first, not the impressive one. If you can't judge the output fast, you can't improve the skill, and you'll blame the tool instead of the task.
- If you're shipping agents to real users, decide now who owns each one and who reviews it. "It works" and "someone's watching it" are two different problems.
- If your agents touch local MCP servers on your laptop, know that endpoint tools are starting to watch that traffic too. Worth knowing before security asks.

What's the most boring, repeatable task in your workflow that you haven't automated yet — and what's stopping you?

## How today's stories score

- **Non-coders at a wealth manager built 24 AI skills:** Signal 46 · Facts 56 · Hype 30
- **Keychron launches a four-key keypad for AI coding assistants:** Signal 80 · Facts 68 · Hype 21
- **Endpoint security tools add visibility into AI agents and MCP servers:** Signal 69 · Facts 67 · Hype 12
- **Open-source Flide reports faster inference for Jev-style models:** Signal 50 · Facts 97 · Hype 29

*Scored by an AI decision model from each story's articles. Facts means how checkable the sources look, not that we verified them.*

## Sources
- [Range's Matt Ragone: Planners With No Coding Skills Built 24 AI Agents in Six Months](https://finance.biggo.com/news/06f2f07f90caf19e) – BigGo Finance, 2026-11-08
- [Keychron launches four-key keypad for AI assistants](https://www.kitguru.net/peripherals/keyboards/joao-silva/keychron-launches-four-key-keypad-for-ai-assistants/) – KitGuru, 2026-11-08
- [MCP Security: Cortex XDR One vs Defender, 50+ Agents [2026]](https://tech-insider.org/mcp-security-cortex-xdr-one-vs-microsoft-defender-ai-agents-2026/) – tech-insider.org, 2026-11-08
- [Flide claims up to 9.5x lower latency for Jev-style decision models](https://news.lavx.hu/article/flide-claims-up-to-9-5x-lower-latency-for-jev-style-decision-models) – news.lavx.hu, 2026-11-09