hartem_

Declarative DSL is a really interesting approach, especially since you’re exposing it directly to the users. There are some applications where throwing the dice in production by having LLM as part of the runtime is not an option.

Can someone please explain why does this work? Did they agree what everyone would be playing? I am not a musician, but it seems like tempos and time signatures do not match. Or is the whole point that they didn't agree to anything upfront and it still somehow works? Are they all actually playing in B-flat?

Not yet because honestly we were able to get to where we needed in terms of accuracy without it. Creating a feedback loop and turning into an agent-style interaction can be helpful for more complex automations, but honestly, you can get a lot of mileage out of what we have. Having said that, agents look really fascinating and some demos I saw are mind blowing, so we will definitely look into it very closely.

Come work at Bardeen! email me artem@ if you're interested

On a more serious note, we have built a DSL and an engine for executing automations inside the browser. Where possible we connect natively to the apps we're integrating with (like Calendar or Notion), where it's not possible (like LinkedIn) we use browser capabilities to interact or get data. We use a GPT model to transform the description of automation to our DSL, we then verify it, typecheck it, fill in the gaps where possible and present it to the user. If the user likes it they can save it and start using it. Happy to share more details if it's interesting.

It's similar to javascript, but not exactly identical. Here is an example that sends a summary of calendar event two minutes before it starts (uses gmail, openai and slack):

``` function (recipient) {\n when: __0 = GoogleCalendar.when_next_event_is_in(time: B.Duration(120000));\n __1 = GoogleCalendar.get_next_event();\n __2 = OpenAI.get_summary_of(text: __1.description);\n __3 = BardeenCommons.get_string_concatenating_strings(strings: ["Your next event is:", __1.summary, "at", __1.startTime, "Here is a summary of the event:", __2]);\n Slack.send_message(message: __3, recipient: recipient);\n}' ```

Initial accuracy was about 10%, which was pretty meh TBH. With a lot of tweaking and tuning we were able to get it to 70%. This means that it takes about 2-3 attempts to get it to generate what's expected.

The great thing is that we only use AI to generate the DSL description of the automation and let the user tweak and tune it. Once it's there we just execute it with our engine.

What was the most interesting thing that you learned while implementing the WAL? Have you thought about how WAL is going to work in the multi-master setup?

Aren’t you using 1Password for generating one time passwords? Or do you deliberately want to keep them out of it?