This is early, evolving work — please jump in with questions and comments at any time. The text on every slide is live and editable.

How can an AI agent join a human pair-programming team?

Orbit connects an AI agent to a running Smalltalk system through the Model Context Protocol. The agent can read, evaluate, compile, test, and refactor live code as a genuine collaborator.

There is no build and no restart. The agent works inside a live object memory, changing the system while it runs — classic Smalltalk livecoding.

An MCP server exposes the image as a rich toolbox: evaluate expressions, compile methods, create classes, run tests, and trace senders and implementors.

A live WebDAV filesystem projects the image as files — /classes, /processes, /search, /sessions. Reading and writing has real side effects in Smalltalk.

Roles and capabilities scope what the agent may do: MethodCompilation, SharedVariableAccess, AgenticMemoryAccess. Permission is explicit, not assumed.

The agent's context is externalized into agentic memory the image itself provides, so understanding of the system accumulates across sessions.

The whole experience is direct and tangible. This deck lives in a morphic-window right beside Caffeine — drag its titlebar, resize it, and edit any slide in place.

evaluate · compile · createClass · addMethod
runTests · getAllSenders · getAllImplementors
traceCallChain · snapshotImage …

Thank you! Questions?