Modern programming languages weren’t built for agents.

Observability and profiling

Agents write more code than any human can read. Telemetry is the only way to understand what happened. BAML traces every function instead of sampling a percentage. It’s 6x faster than OpenTelemetry in Rust, 200x faster in Python, and the traces 1000x smaller.

AI functions

Calling a model should feel like calling a function, not wiring up an SDK. In BAML it’s a typed function: declare the input and output types, get structured data back. Malformed output gets repaired against your types by schema-aligned parsing, so every model gets better at structured output.

The anti-slop type system

Every invariant you can’t enforce is one an agent will eventually violate. Start with type erasure: we don’t do it, so there’s no any or unchecked cast for a model to hide behind. In BAML, invalid states don’t compile.

interface User {
  name: string;
  email: string;
}
// 'as' is an unchecked promise the compiler believes
const user = JSON.parse(raw) as User;
user.email.toLowerCase();

Agent-first toolchain

For decades, tooling was built for humans: LSPs, autocomplete, hover docs. It’s about time the real author of the code got fair treatment. An LSP for you, baml describe and friends for them. And a really fast compiler for both.

$ baml run main
   Compiling 1 file(s)
    Compiled 1 file(s) in 1s
"hi, world"
$ baml run greet -- --name "hacker news"
Greeting { message: "hi, hacker news" }
$ baml run greet -- --help
function greet(name: string) -> Greeting
Options:
      --name <string>
# any function is a target; its params become --flags

Workflow primitives

How has everyone accepted async as a good idea? Half your codebase ends up as two copies of the same function, one sync and one async. Like Go, BAML has green threads. Unlike Go, you can await what they return.

Better testing

Engineers spent twenty years squashing flaky tests. Then AI made everything flaky. BAML tests can grade distributions. Cases can be hard-coded examples, golden datasets, or real prod traces.

testset "from object storage" {
  let res = baml.http.fetch("https://datasets.example.com/golden.csv");
  let rows = res.text().split("\n");
  for (let row in rows.slice(1, rows.length())) {
    let cols = row.split(",");
    test ("golden: " + cols[0]) {
      assert.equal(classify(cols[0]), cols[1]);
    }
  }
}

I’m hyped. How do I port my codebase to BAML?

Do not rewrite your codebase in BAML.

You can if you want. But we went the extra mile in the other direction: every type, every function, every method crosses the bridge to your language. Even generics. Even lambdas. Even tracing. Sht just works. Read more → Incremental adoption

Can models actually write BAML?

Decide for yourself. BAML reads like TypeScript and is already an official language on GitHub, so models should know most of it. We don’t leave the rest up to vibes. We measure how agents write programs: how long it takes, what it costs, how many turns. baml describe wasn’t a coincidence. It’s a science.

And remember, today is the worst the models will ever be at BAML. Read more → Agent tries BAML

Is there an ecosystem?

npm install is-even. It depends on is-odd? That depends on is-number?! 🤯 We’re rethinking package management from first principles. For now, we’ve shipped a thorough standard library. For anything else, ask Claude to port what you need, or pass functions over the bridge.

Besides, no packages means no supply chain attacks. QED.

Are you trying to get bought by Microsoft?Anthropic?

No thank you.

The BAML language is and will always be open. Apache-2, free, works offline.

We make money on Boundary Web Services. Build the language, the runtime, and the tracing layer, and you can make things nobody else can. We think you’ll love paying for some of them.

Some of you may wish to build your own cloud. Good luck and Claudespeed.

For the rest, we’re launching with observability, and saving you from Datadog. Reach out if you want in early.