Internal tool
Orchex.
Built in the studio, now runs inside it.
A parallel AI orchestration engine the studio built from a whiteboard idea. It splits a piece of software work into streams, runs them across several language models at once, and checks the result. Today it runs inside the studio’s own work.
THE PROBLEM
AI assistants are brilliant soloists. They just can't work in teams.
Serial. Fragile. Forgetful.
- › One task at a time, even when tasks are independent.
- › Rate limits kill momentum after 20 minutes.
- › Two agents on the same file silently overwrite each other.
- › Any failure, and the whole run starts from zero.
Parallel. Owned. Resumable.
- › Independent tasks run in parallel waves.
- › Seven LLM providers, smart routing around rate limits.
- › Ownership enforcement: each stream declares its files. Streams do not collide by design.
- › Failures self-heal; successful streams persist.
None of these are opinions. They're the daily lived experience of everyone building software with AI in 2026.
ARCHITECTURE
One system diagram. Four decisions we made early.
Orchex Engine
Agents produce text, only.
All file I/O lives in the orchestrator. No permission battles, no tool-use friction. Runs inside the host the studio already uses.
Ownership enforcement.
Every stream declares which files it can modify. Violations rejected before disk write. No merge conflicts by design.
Provider-agnostic from day one.
Seven LLMs behind one interface. Switch providers in one line. Route cost-sensitive vs reasoning.
Self-improving, visibly.
Every run generates an execution report. Patterns surface after 5–10 runs. Learning is shown, not hidden.
LESSONS
Three things we'd take into any client project.
Don't fight the platform. Redesign around it.
We spent two months building v1 through v4 of orchestra. All of them broken, because we tried to make agents do what the platform didn't allow. v5 worked because we stopped fighting.
Constrain first. Generalize later.
Ownership enforcement is a hard "no," and that is exactly what makes the rest safe. Every strong system has one or two ruthless constraints doing the heavy lifting.
Ship the ugly version in weeks, not the beautiful version in months.
Orchex v3.0 shipped in a week. It was ugly. It worked. Every version since has been polish on a thing the studio was already using. Beautiful plans die in planning.