Scout

Scout is a tool I build at Cascade, where I’m a co-founder, for teams that ship code with AI agents every day. It keeps one record of what your team decided, and every agent reads it before it starts work, so your agents stop building on decisions you’ve already changed.

If you run more than one coding agent, you’ve probably watched one of them confidently build something the team walked away from last week. The agent wasn’t wrong about anything it could see, because nobody had told it.

Visit runscout.dev →

Why AI agents work from outdated decisions

An AI coding agent starts every session with only what it’s handed. Usually that’s the repository, a ticket, and whatever rules file the team keeps, like a CLAUDE.md or a set of Cursor rules.

The real decisions usually get made in the middle of a session, when someone drops a feature to hit a date, or decides one customer gets handled differently, or changes how the product behaves because the old way was wrong. That happens inside a model session, or in a Slack thread, or in someone’s head while they read the output. The ticket was written before any of it, and almost nobody goes back to update the ticket or the rules file. I wrote about that gap in The Handoff Problem.

So the next agent, maybe in a different tool, run by a different person, starts from the old version. It does good work on the wrong plan, and someone finds out in review, or later.

How Scout works

Scout connects to your agents over MCP, and when a session starts, the agent makes one call and gets back the current state of things:

  • What changed since it last checked in.
  • The team’s rules.
  • Patterns the team has stopped using.
  • Work a teammate passed along.
  • What each teammate is working on.

Every agent gets Scout’s instructions when it connects, and those instructions tell it to read the record at the start of each session and to write back as it works: a decision, a lesson, a plan or a blocker, each with who made it and when. Nothing gets deleted: when something changes, the new entry wins, and when two entries disagree, Scout flags it so a person can settle it and say why.

Setup is one MCP connection, with guides for Claude Code and Cursor, and it works with any agent that speaks MCP. Scout itself is a hosted service on Fly.io, so there’s nothing to install or run on your side. The security page covers how it handles your data.

What it costs

Current pricing, the trial and the startup program are on runscout.dev/price.

Who it’s for

Scout is for tech leads, CTOs and founders of small teams where several people, each running their own agents, work on the same codebase. If you’ve ever had to explain the same decision to three different agents in one week, or found an agent rebuilding something you’d already ruled out, this is the problem it’s built for.

It’s not a coding agent, a wiki, or a chatbot for your repository, and it doesn’t summarize your history. It only knows what someone, or some agent, writes to it, so a decision that stays in a Slack thread is still invisible. And it only helps agents that speak MCP.

Why I built it

Most days I have more than one coding agent working on the same codebase, and I kept hitting the same thing. An agent would do clean, careful work on a plan we’d changed two days earlier. The new decision was real, but it lived in a session that had ended, or in my head, and the next agent never saw it.

A rules file helps, and for a small, disciplined team it can be enough. It stops being enough once several people and several tools are moving fast, because going back to update it is the step that gets skipped. Scout puts that step where the work happens: the agent writes the decision down when it’s made, and every agent reads it at the start of the next session. We’re building Scout with Scout, so every one of our own sessions starts from it too. Most of what I write here is about judgment, because when agents do more of the building, the decisions are the part people still have to get right. Scout is how I keep those decisions in front of my own agents.

I write about judgment and AI every Tuesday in my newsletter, and you can subscribe below.

Start with Scout at runscout.dev →

Frequently Asked Questions

What is Scout?

Scout is one record of what a team decided, changed and learned. Every AI coding agent on the team reads it when a session starts and writes to it as it works. It runs at runscout.dev. It's not ScoutOS or Scout APM, which are unrelated products with the same name.

Who built Scout?

I did, with the team at Cascade AI Partners, where I'm a co-founder. We use it every day on our own work. We're building Scout with Scout.

How is Scout different from a CLAUDE.md or Cursor rules file?

A rules file is written by a person, usually after the fact, and it lives in one repository for one tool. It's a good place for rules that don't change. Scout is where the decisions that do change go: agents write them down as they work, the newest entry wins, and when two entries disagree Scout flags it for a person to settle. Every agent on the team reads the current version at the start of its next session.

Which tools does Scout work with?

Any agent that speaks MCP. There are setup guides for Claude Code and Cursor, and in Claude Desktop or claude.ai you add Scout as a connector and sign in.

What does Scout cost?

Current pricing, the trial and the startup program are on runscout.dev/price.

Facing a technology decision? Let’s talk it through. Get in touch →

Colorado Front Range · In-person and remote · Schedule a call