Stop using browser agents.
@tanishqk1.7K followers
56,382 views
45 Comments
Comments from the original post on X (Twitter)

@tanishqk
"what happens when the dom moves?" - libretto prefers making internal API requests where safe & secure. So that makes it even more reliable and simultaneously way faster - for the integrations with legacy portals that we're focused on, the DOM never moves. Those companies are just 90% sales people. - when it does change, you should be in the loop and own that change. For important business-critical automations like filling out tax forms or in healthcare, you want a human reviewing and owning that change. Still, for less critical cases, we'll have an auto-review agent mode soon! Gets 90% of the benefits.
4 likes

@tanishqk
also here is the account to follow @libretto_sh
1 like

@ccorcos
@Neesh774 Next stop, Hollywood!
3 likes

@tanishqk
@Neesh774 available to hire as an actor
3 likes

@YigitTabel
Checking the "public-website-benchmarks.md" file in your Github repo. I saw this benchmark. I got that this is the structure you have in mind: "build automation once → replay cheaply → detect breakage → repair with AI → update automation". My skepticism: For hostile or highly dynamic public websites, Libretto-cached seems fragile to me. For semi-stable authenticated portals, internal dashboards, admin panels, insurance/healthcare portals, vendor portals, and repetitive B2B workflows, it can be very useful. Did you manage libretto to overcome browser security challenges and the other workflow-stopping browser automation blockers?
3 likes

@tanishqk
that's fair and thank you for digging into the details! We wanted the benchmark to be easily auditable and reproducible by other people hence the focus on public no-login websites. Libretto workflows by default can solve some security challenges by themselves - if the agents can go through them, they will put it into the workflow code. For more production captchas (Cloudflare, Akamai, Datadome), we recommend running on a cloud platform (like Libretto Cloud, but also Kernel, Steel, Browserbase) that have auto captcha-solving built in. The use-cases you laid out re: admin panels, insurance/healthcare portals, vendor portals, and repetitive B2B workflows is exactly the kinds of workflows we want to be supporting Does that help answer your question?

@jia_seed
LFG!! 👑👑 proud of the whole team and this looks awesome
2 likes

@tanishqk
thank you!!
1 like

@sam_wise_
This is the right split. Use the coding agent to generate the boring deterministic path. Save the browser agent for the genuinely ambiguous steps. Otherwise you pay tokens to simulate a script poorly.
2 likes

@tanishqk
completely agree!
1 like

@tanishqk
If you want to convert your browser agent to fast, cheap, deterministic code, here's my scheduling link
2 likes

@rsms
Congrats on launching!
2 likes

@tanishqk
thank you rasmus!

@__morse
super cool
2 likes

@tanishqk
Thanks Tommy! Thank you for building Playwriter, the original source of inspiration
1 like

@KyranTanDev
Super sick! I’m curious about Libretto’s error handling: e.g. if a form throws an error after a field edit, does the workflow recover, or halt? I’ve been building in this space myself (a computer-use framework that turns repetitive tasks into deterministic automations to cut token cost) and I’m really keen to see how you guys approach this.
1 like

@tanishqk
good q. It's just code - by default, if the agent only writes code for the happy path, the workflow will just fail and you'll see the error. When that happens you'll get an email and you'll actually be able to connect to that running session that ran into the error. Once the agent connects to it, it will figure out how to work around the error, you'll be in the loop, it fixes and deploys the updated workflow.

@GaddipatiHarsha
Let’s goo
1 like

@tanishqk
thanks man!

@realpraty
Headless interactions are the future
1 like















