HN user

nashadelic

235 karma
Posts4
Comments142
View on HN

Compiled agents: http://squig.com/

It takes your instructions, write a versioned spec, then generates a hybrid workflow of code+LLM calls and protects it with tests/evals

The result is that the agents run much faster (90% of it is code), cheaper (LLM steps are scoped tightly and uses smaller models) and reliably (specs get turned into coded state-machine)

Claude Fable 5 1 month ago

I've never engaged with their support (I have dedicated POC), but they don't use AI for their support?

Claude Opus 4.8 2 months ago

The only real way to see this if you have consistent evals for common usecases in your B2B SAAS product and see if the tricky usecases are being solved. You'd then go down to the cheapest model that can solve the evals.

Dropbox has deep integration ecosystems, runs your company's data, I think its a no-brainer for it to become the agentic memory for your company if done right with it syncing data across all company services.

I've been thinking about this and I believe the best place to be is a scientist who keeps looking at an AI's output, prods it in the right directions, verifies the proofs, fixes and fills gaps, takes the proof to production with safety, risks etc mitigated and then distribution with a company wrapped around the discovery. I think it wouldn't be black-boxed as much and will require a lot more understanding and reviewing to trust and productize it.

Whats funny is that nearly every major desktop app and feature has come out on macOS first and windows has nearly always been an after thought

I'm very curious about how much of it is re-using (remixing?) an existing browser implementation it has seen in the while or its trained on. Even in that case, 99% of all code isn't doing anything novel so this copying would still have significant practical use.