Orchestration Solution (after 6+ months of analysis paralysis) Wrote a declarative graph executor DSL in TS that generates DOT diagrams and compiles into Restate durable workflows Green:

BAML

functions etc. (running in Restate service handlers) Orange: Restate Virtual Object (Representing the conversation,

appendTokens

while streaming,

commitMessage

for when a generation completes,

getState

FSM in a union) Cyan: Durable Restate Promises https://cdn.discordapp.com/attachments/1119375594984050779/1497675982293827694/image.png?ex=69ee62f3&is=69ed1173&hm=a9fd0261180890c66dd02d2ad4bcac3a25ddee9d9b267e68804c4d69d9faff4d&

I originally started using restate in the project after seeing your considerations

Wait really? Like long ago when I wrote something about it in here?

Yes, at the time I was considering Vercel's workflow and Restate

That's why I ended up designing this "DSL". I realized the biggest blocker for me often was about being able to clearly define use-cases, and how systems depend on each other etc. I used Claude Code to design the API, it basically broke down Restate into a more simplified higher-level API (also less capable), but I've been surprised with how well Claude seems to understand Restate and how things should compose etc. My plan now is to define as many systems I can using that DSL, then when I hit limitations where I need imperative Restate code, I'll try making Claude do it for me... Damn I'm such a weak dev 😭

After you shared your design, I immediately asked Codex GPT-5.5 to analyze its strengths and the complexity of my own implementation to see whether there was anything worth improving. Codex said your design is relatively clear, but much of my current implementation still can’t be directly integrated.

Yeah, Claude also need a little bit of whipping to realize how to use the DSL to convert natural language into programs... hehe Spent a long time on my docs to make it work well! And when I hit limitations.. guess what I'll do. That's right. I'll ask Claude to expand the feature set of the DSL. Nothing could go wrong, right?

I remember when I was writing the design document for restate + BAML, I spent a week on it before I dared to start coding、

What are you building? Just out of curiosity

I shared it in the community; here’s the link https://discord.com/channels/1119368998161752075/1253172394345107466/1462994235203522723 https://discord.com/channels/1119368998161752075/1463061149070393505/1463062812199882794 Overall, it is a general-purpose orchestration agent, plus Bash tools

Interesting, so you're just testing stuff out and learning?

It’s not just for testing—it’s already been applied in my product, even though the product isn’t great and not many people use it. BAML’s philosophy has always appealed to me, but BAML’s new features have indeed taken far too long.

What is your product? 👀

An AI assistant for the automotive industry

This is how my DSL would implement a map reducer btw:

typescript
const carry = flow.pointer()

const mapper = flow.add('Mapper', mapperDef)
  .dependsOn(carry)
  .if('Has Carry', r => r.resolved(carry))
  .else(flow.input)

const reduce = flow.maps(mapper).via('Reduce', reducerDef, 'reduce')

const flatMap = flow.add('FlatMap', flatMapDef).dependsOn(reduce)
  .loop(carry)
  .if('Iterate', pred)

flow.maps(listProducer).via(callback)

is the fanout mechanism, representing "Spawning N dynamic nodes to process a list" (it's all separate Restate invocations) Of course if you don't want separate Restate invocations, you could just implement it internally of a service handler that accept a list as it's input... But that would be if it's not a semantic concern of "orchestration logic" and strictly belongs to business logic

Another brand-new concept—let me first have codex explain it to me

Yeah the elegance is tempting right