← Work

Tooling

Orchestra

A minimal-input, spec-driven pipeline that orchestrates AI agents to build a whole product — software, website, research, and pitch deck — from one brief to shipped.

Architecture diagram: one spec fans out into parallel arms — build pipeline, web, research, and pitch deck — that converge into a shipped product, with a continuous audit daemon running underneath.

Problem

Building software with AI agents by hand doesn’t scale. You babysit every step, the context lives in throwaway chat windows, and nothing actually takes an idea and finishes it.

Approach

Orchestra is a spec-driven factory. Give it one brief and it coordinates a fleet of AI agents to produce a whole product end to end — running parallel arms for the software build, the website, market research, and a pitch deck. A persistent operator coordinates; ephemeral builder agents do the work; they hand off through flag files on disk, and every artifact is committed to git. Minimal input in, a shipped product out.

Diagram comparing two task-dispatch methods: a linear A to B to C handoff chain, versus an operator that keeps the work open and spawns fresh builder agents iteratively with an ad-hoc grader deciding when it is good enough
How the build loop works — a linear handoff chain versus the operator + fresh-builder loop the system uses.

The hard part

Each arm runs its own operator-and-builder loop with an audit-redo-grade check, and a background daemon audits the whole codebase on a schedule — rotating a five-model pool every few hours with a daily triage pass, then committing its findings. Continuous, multi-model review, running as infrastructure.

A handwritten notebook page mapping the build pipeline stages
Where it started — the pipeline stages, mapped by hand.

Outcome

More than 1,600 commits; it now drives my own development across the projects on this site.