A meta-harness for AI coding agents. Not another coding agent, and not a wrapper around one - the layer above them that decides what to build, hands each piece to the right agent, checks the result against a contract, and keeps a durable record of what actually happened.
Coding agents are good at doing a task. They are bad at knowing which task, keeping a week-long goal in view, noticing they've drifted, and telling you honestly whether the thing is done. Skippy owns that outer loop.
A conversation becomes a brief. A brief becomes goals. Goals become features, features become tasks, and every task is judged against acceptance criteria it was handed up front. Understand, plan, build, verify - with a human pulled in only when a real decision is needed.
I argued in We've Been Here Before that agentic AI is rediscovering the hard lessons of distributed systems - tracing, verification, circuit breakers, observability. Skippy is me putting my money where my mouth is: a harness that builds those in from the start. It's the project I'll be writing the most about.
A closer look at the surfaces Skippy runs on. Click any screen to open it full size.
The name is a tip of the hat to Skippy the Magnificent - the wildly overconfident (and secretly indispensable) ancient AI from Craig Alanson's Expeditionary Force series. My Skippy is considerably more humble and does not, as far as I know, insult anyone. Full credit to Alanson for the character and the inspiration.