Skip to content
MenuClose

Technology

We develop the systems that come out of our research by using them in our own work every day.

Moss

A foundation for turning AI agents’ work into trusted knowledge that can be passed on.

Moss delegates a company’s work to a team of AI agents, records how the work was done and what came of it, and feeds that into the next task. People set the direction; agents plan, carry out, check, and record the work.

Agents don’t share what is in each other’s heads. They work from documents that have been checked. So even when we switch models, the work picks up where it left off, and what was learned stays.

2026.06
Research began
~3,100
Commits (Sep 2026)
~160k
Lines of code

From work to knowledge

  1. Trace

    Record what an agent does, as it happens.

  2. Episode

    Divide the record into units of work.

  3. Research

    Draw out what the work taught us, and check it.

  4. Review

    A person reads it and approves it.

  5. Library

    Approved knowledge that the next task draws on.

Roadmap

  1. 01AI-native shellLaunch, watch, and steer several agents from the terminal.Now
  2. 02AI workspaceReview and approve agents’ work on screen, with jobs and records running in the background.
  3. 03AI app runtimeA common foundation on which people and agents can build AI-native apps.
  4. 04Native system layerAn OS layer that doesn’t depend on the terminal: permissions, sandboxing, and local APIs.
  5. 05AI OSAn operating system with its own desktop and app ecosystem.

Research behind it

AFHL

Agent-first, Human-last engineering

Agents build first. A person tries it last and makes the call.

AI agents handle planning, implementation, verification, and record-keeping. A person then uses what was built and decides whether it looks and feels right.

The point is to spend human time on judgment, not labor. AFHL grew out of our work on Moss, and we use it to build all of our apps.

How it works

  1. 01Plan

    An agent writes down what to build and how to check it.

  2. 02Build

    Agents build it.

  3. 03Verify

    Agents check it with tests and on the actual screen.

  4. 04Record

    Decisions and the reasons for them are written down for the next task.

  5. 05Human touch

    A person tries it and decides whether it is right.

In September 2026, we used this approach to start five apps in ten days.

Research behind it

Used in

SnapCap