Kyle Jeong
@kylejeong
649 Following 8.5K Followers
what is actually going on with computer use right now?
i'll be on @MTSlive at 12:30pm to talk about the state of computer use and what it actually looks like to deploy it to production.
there is beauty everywhere for people with eyes to see
the ability to write really good internal memos is underrated
i confirmed instinct was AGI by having it figure out how to get me out of jury duty
How do you actually build an effective harness with Claude?
We had Thariq (@trq212) from Anthropic at Navigate 2026 to talk about "Unhobbling Claude", the difficulties and processes on how to build agents and harnesses.
Show more
there's 2 sides to this, bot providers and anti-bot providers.
anti-bot is meant to defend websites from malicious traffic, whereas bot providers have to verify the traffic they're enabling is not malicious
every single agent will need to verify themselves to prove their intentions on the internet.
Show more
This is a big enough problem that there's not just one startup idea here, but several. For example, there will probably be one variant meant to filter out agents, but another meant to welcome them in an orderly way. This also seems an idea you could build more on.
Show more
i'm extremely curious as to how people outside of the valley are adopting muse
it’s now the #1# most popular app in the app store over ChatGPT.
are normal people actually using muse?
Show more
alexandr wang's poasts are his version of the "zuck glow-up". he's becoming "cool" again by just not caring what people think and being himself online.
a nightmare to comms teams across the globe, but honestly working very well for muse's distribution.
Show more
if you'd like to test this, shoot me a DM!
I built JevSearch, search the web & validate your results with Jev.
Give a query and selection criteria, use @browserbase search to get the t25 results, then Jev scores and returns the t5 results.
Jev often chooses urls outside of the initial top 5 as more relevant.
Show more
Stripe's new WebMCP endpoint lets your agents make full purchases on the web.
Using @Stagehanddev you can navigate to any Stripe checkout page and use their new WebMCP tools to fill the checkout and fetch order details.
Show more
posted to our blog if you prefer to read there:
This is experimental, so we'd love feedback!
Checkout this branch to try it out:
Extract is especially interesting because Jev can’t generate text.
Instead, Jev picks the elements containing the values. Stagehand copies their text, parses numbers, and resolves URLs.
For repeated rows or cards, Jev identifies the fields in one item.
Show more
Observe uses the same selection process.
Looking for one element? Jev picks a candidate.
Looking for several? We ask whether each candidate matches, batching the questions together.
Stagehand returns the matching elements as executable actions.
Show more
For Act, we break the instruction into smaller decisions. “Type ‘San Francisco’ into the destination field” turns into:
Identify the action: fill
Identify the text: San Francisco
Choose the destination input
Then deterministic code fills the field and checks the value.
Show more
Most browser actions involve choosing among things already on the page.
Stagehand turns the page into candidates, with context like labels, headings, and table rows. Jev picks the action and Stagehand executes it.
If the choice is uncertain, we fall back to an LLM.
Show more
In early evals, we're seeing adding Jev cut median Act latency from 1.97s to 0.46s (about 4.3× faster)
Across 240 real-site task runs, success went from 85% to 95.4%.
Jev wasn't meant to build Agents.
Instead, we added it to @Stagehanddev's AI-powered primitives Act, Extract, and Observe.
Rather than a fully autonomous agent, we use Jev as a decision layer: which element to click, field to fill, or text to extract.
Show more
Jev-pilled son or fly brain neural net daughter
