Design decisions, release stories, and longer-form essays on building a language for auditable AI orchestration. Written by the people building it.
We're publishing the first alpha of Cleat, a statically typed orchestration language designed for the agentic era. Here's why we built a new language instead of another framework, what's in the box today, and where we're going from here.
needs is a keyword, not a decorator
Decorators give you the syntax of effect tracking without the enforcement. We made needs part of the type system from day one. Here's what we tried first, what broke, and what convinced us to bake it into the grammar.
Good error messages are a feature, not an afterthought. A walkthrough of how Cleat's diagnostics work — from token spans to the squiggly underlines — and why we copied Rust's approach instead of Go's.
Every AI orchestration framework eventually adds a "logs" tab. They are all insufficient for the same reasons. An argument for why cryptographically signed, append-only chains belong in the language — not bolted on as a SaaS add-on after the fact.
A guided walkthrough of building a real, useful tool with Cleat — from tool declaration to mocked tests, with every line explained. No toy examples. No "imagine you have an LLM" handwaving.
When function A calls function B which calls function C, who declares the effect? We picked one answer; the other answer is also defensible. Notes on the tradeoffs and why we landed where we did.
The Cleat type checker rejects unreachable states and detects deadlocks at compile time. The trick is older than you think — TLA+ has been doing it for thirty years. A short essay on standing on giants.
The first post. Why we didn't write another Python library, why we didn't fork an existing language, and what convinced us that orchestration deserves to be a first-class linguistic concern. The mission statement, before there was any code to back it up.
Roughly one essay a month. Design decisions, release notes, the occasional postmortem. No spam, no marketing.