Selected work / Project profile
Constellation
Getting several AI agents to work on a program raises a practical question: how do you coordinate their contributions and check the result?
I authored Constellation to organize that work across Claude Code, OpenCode, and Codex. It is my flagship project, using file-based orchestration to turn objectives into tested programs.
From an objective to a checked result
A coding agent can produce code and report that it is finished. Accepting the work takes another step: checking whether the program behaves as requested. With several agents involved, their contributions also need to reach the right place with enough context to be understood.
Constellation brings coordination and verification into the same workflow, rather than leaving the work scattered across separate conversations.
How it works
The architecture uses model dispatch, file queues, and verification gates. Dispatch routes work to a model. Queues hold work in files. Gates put checks between producing work and accepting it as complete.
Files give the coordination a place outside the chat itself. The system works across Claude Code, OpenCode, and Codex and runs both locally and on a virtual private server, or VPS.
What I'm working on
Constellation is an ongoing focus of my AI work. Its combination of file-based coordination and verification reflects my preference for systems that leave something to inspect: the requested work, the output, and reasons to accept it.
A passing check answers a particular question about a program. Deciding which questions matter is still part of the job. I explore that distinction in Verification before done, with practical suggestions for planning checks and handling incomplete results.