The assistant that speaks first

Every AI assistant I own is waiting for me to tell it what to do. It sits behind a blinking cursor, patient as a bartender at three in the afternoon, and does precisely nothing until I think of the right question. That’s a tool. A good one, mind you… but a tool.

I wanted an executive assistant, a colleague, a collaborator. The kind who leans over after a meeting and says, “You’ve asked me the same thing four times this month. Want me to just build it?”

In KCS terms, every assistant on the market lives in the Solve Loop. A request comes in, an answer goes out, everyone feels productive. Nobody is minding the Evolve Loop: the part where you step back, look at the pattern across all those requests, and fix whatever keeps generating them.

Scout is my attempt at an Evolve Loop for one person. It’s a predictive collaborator rather than a subordinate task machine. It reads the sources I connect (my Claude Code sessions, my git repositories, my calendar, the email threads I’ve replied in, the Notion pages I choose), notices patterns, friction, and open loops, and speaks first.

The repository’s first commit landed at 11:15 a.m. on October 17. It was documentation: platform notes, design decisions, architecture. No code. I’ve spent enough of my career cleaning up knowledge bases built without a taxonomy to know where the time actually goes.

Then it moved quickly. By 12:28 the command-line version produced its first digest from my real data. The menu bar app was up at 2:34. Gmail and Notion were connected by 3:04. Phase 2 started at 5:35 with Scout’s own MCP server, so Claude Code can ask Scout what it has noticed. By 8:09 the database was encrypted at rest, with its key in my Keychain, and Scout could run its own models without LM Studio.

To be plain about it, Claude Code wrote most of the code. My job was the job I do at work: set the constraints, make the calls, write down why, and say no. Every decision went into a file called SCOUT.md with its reasoning attached… thirty-eight of them, and counting. That file is the real product of the day. Code is cheap now; reasons aren’t.

A few of those decisions are worth the price of admission:

Most of the engineering went into column three of that diagram: the gates that hold a pattern back before any model writes a word. An assistant that speaks first is mostly defined by what it doesn’t say.

It also builds as two editions from one codebase: a direct one I’m dogfooding now, and a sandboxed App Store edition with the seams already in place. Whether anyone else ever gets it depends on whether it earns its keep with me first. I’ve spent a career telling organizations to prove the value before they scale the program; I’d be a hypocrite to skip that step with my own toy.

The bigger lesson isn’t about Swift or models. Knowledge management is quietly becoming context management, and Scout is what that looks like at the scale of one person: the value isn’t in the answer, it’s in noticing which question you keep asking. The assistants we have are very good at the Solve Loop. The interesting work, for people and for software, is in the Evolve Loop.

So the question I’m sitting with is this: what would your week look like if something read it back to you, every morning, and only spoke up when it had a point?

Transparency note: drafted with Claude from the Scout repository and its decision log, then read, cut, edited heavily, rewritten, and finally approved by me. As always: AI drafts, humans ship.