sgtwompwomp

This is dope, is this kind of like Wafer.ai but for local models? As in a coding agent optimizes the kernels so the local model runs continuously better? Cause that is compelling if so. If it’s more simple that’s cool too

Awesome work. If I understand correctly, your service essentially runs a bunch of web agents to maintain this dataset and keep it up to date and accurate, and we access through your API / MCP?

You can try https://www.vals.ai/vals-smith for this, saw it recently.

It's not a business's job to make their documentation accessible to their potential and current customers?

I would ask if you've started a real business but it's clear you haven't. It is 100% on a developer tool startup to provide documentation that is easily accessible. If they don't, customers will struggle to get value. If you think this isn't true, then you are ignoring the gigantic market of companies purchasing documentation products (look at Mintlify's customer base for reference)

There is no way I'm asking my customers to scrape my docs and build their own MCP server and AI assistant just to access it easily.

Mintlify is really good. If you're a serious developer tool not sure why you wouldn't use it. For example I went into your docs and I don't see AI chat so I can ask quick, natural language questions. No MCP I can install so Cursor can query. Prob no llms.txt. No quick Copy to Markdown. This stuff is table stakes, if you don't have it and a competitor does, I'm not even considering you guys

It's just a worse developer experience. Fine if you aren't a serious business, but yeah I wouldn't play down the value of Mintlify or similar products. It's seriously good and it's why huge companies use it

Yup, this is a case where you always want an agent to do that step. So in the prompt you just say “do a focused_action to select the search result with John”, and then the pathfinder agent will cache in it’s memory to delegate that step to a mini computer use agent, just for that particular task.

After the focused action is done, it’ll go right back to deterministic!

Our agent would have a tool to essentially bring in the human. Not built this yet, but the closest thing we do have is that our agent can declare a task as failed if it determines it can’t proceed (based on your instructions).

More on this soon! How would you imagine this would be useful?

So once you run a workflow (prompt) once, then that trajectory is cached.

Run the workflow again and it’ll run through that cached trajectory as best as it can, falling back to computer use if needed.

So what I meant is this: When you run our Cyberdesk agent the first time, it runs with the computer use agent. But then once that’s complete, we cache every exact step it took to successfully complete that task (every click, type, scroll) and then simply replay that the next time.

But during that replayed action, we do bring in smaller LLMs to just keep in check to see if anything unexpected happened (like a popup). If so, we fall back to computer use to take it home.

Does that make sense? At the end of the day, our agent compiles down to Pyautogui, with smart fallback to the agent if needed.

Unfortunately these scripting tools just are untenable when dealing with so many desktop flows that all have changing UIs and random popups. You end up having to repair all of them all the time, in fact there's a whole consulting industry out there just to do this all day.

The whole idea of Cyberdesk is the prompt is the source of truth, and then once you learn a task once via CUA, the system follows that cache most of the time until you have to fall back to CUA, which follows the prompt. And that anomaly is also cached too.

So over time, the system just learns, and gets cheaper and faster.

Yup, a few of our clients have a need to verify something in the software, so we support an agentic step where we look at the screen and can verify whether something exists, or whatever a step was completed, etc!

Thanks! And yes, so our pathfinder agents utilize Sonnet 4's precise coordinate generation capabilities. You give it a screenshot, give it a task, and it can output exact coordinates of where to click on an input field, for example.

And yes we've found the computer use models are quite reliable.

Great questions on scale: the whole way we designed our engine is that in the happy path, we actually use very little LLMs. The agent runs deterministically, only checking at various critical spots if anomalies occurred (if it does, we fallback to computer use to take it home). If not, our system can complete an entire task end to end, on the order of less than $0.0001.

So it's a hybrid system at the end of the day. This results in really low costs at scale, as well as speed and reliability improvements (since in the happy path, we run exactly what has worked before).

There isn't a viable computer use model that can be ran locally yet unfortunately. Am extremely excited for the day that happens though. Essentially the key capability that makes a model a computer use model is precise coordinate generation.

So if you come across a local model that can do that well, let us know! We're also keeping a close watch.

Yup! We have signed certificates that explicitly state this, with all LLM providers we use.

It's on our production AKS cluster :) So fully scalable.

By the way, we've launched on ProductHunt btw! If you're interested in giving us an upvote (same for anyone else here!)

https://www.producthunt.com/posts/cyberdesk

I hard agree. Seth, Ayo, I think you guys should be honest with yourself and ask whether you want to go toe to toe with Cursor, Windsurf, Microsoft, etc.

If not, take the other route. Go deep into the vertical of React Native. Help people with no experience run an ENTIRE BUSINESS just with your chatbot as the AI cofounder - you build the entire app and backend, handle marketing, publish it, all with AI agents. How sick would that be.