SShortSingh.
Back to feed

Yutori's Dhruv Batra Says Websites Should Be Built for Browser Agents That Act Like People

0
·3 views

Dhruv Batra, co-founder and chief scientist of Yutori, argues that most websites will never offer APIs for AI agents, since infrastructure changes slowly. He says large firms holding enterprise and customer data have the incentive to build agent pathways, while the long tail of smaller sites, such as school district and government pages, probably will not. Batra proposes that browser agents perceive screens and click buttons, making the visible interface itself the operating layer. To support this, he recommends using real buttons and links, stable visible labels, and clear on-screen results, so agents can complete actions like checking shipping costs.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Caltech professor stops grading homework, citing widespread AI use by students

A Caltech professor has decided to no longer grade homework assignments because students are widely using AI tools to complete them. The change reflects growing concern among university instructors that homework scores no longer reliably measure individual understanding. The details come from a Business Insider article linked from a Hacker News post, which had a small amount of early engagement at the time it was shared. The article is part of wider coverage of how AI is affecting math classes and how students may be preparing for careers. Further specifics about the course and the professor's alternative assessment methods were not available in the provided material.

0
ProgrammingDEV Community ·

PulseVerse Turns Webcam-Detected Heartbeat and Breathing into Interactive Visual Worlds

PulseVerse is a privacy-conscious experience that uses an ordinary webcam and the Presage SmartSpectra SDK to sense heartbeat and breathing without wearable devices. Instead of showing clinical charts, it renders these signals as 3D star fields, a breathing garden, and other interactive visuals. Heartbeats drive gravitational shockwaves through a starfield, while respiration expands and contracts a bioluminescent grove across four selectable biomes. Additional modes include guided breathing exercises, a constellation generator for exporting artwork without biometric data, and a stress visualization that settles as heart rate variability stabilizes. The project is presented as a self-exploration and biofeedback tool, with sound remaining optional.

0
ProgrammingDEV Community ·

Pandas explode() turns list-valued stock tags into a lookup table and co-occurrence matrix

A DEV Community tutorial explains how pandas explode() can convert a column of tag lists, such as stock-to-concept mappings, into one row per tag. It notes that real-world data often mixes lists, unparsed delimited strings, duplicates, empty lists and missing values, each of which behaves differently under explode. The author recommends first normalising every cell into a clean Python list, then exploding with a fresh index and cleaning up the result. The resulting long table supports tag lookups, tag counts and a co-occurrence matrix showing which concepts appear together. The examples use pandas 3.1 with placeholder tickers and company names.

0
ProgrammingDEV Community ·

Study Finds Flawed Telemetry Can Mislead AI Agents Into Wrong Health Conclusions

A joint post by OllyGarden and Causely describes an incident where an agent queried logs for ERROR-level records, found none, and wrongly reported a payment service as healthy. The severity was embedded in message text rather than the severity field, so the query returned an empty result that looked valid. The authors argue that telemetry can be well-formed and schema-compliant yet still fail to reflect what actually happened, which is harder to detect than obvious problems like log storms. Drawing on audits of OpenTelemetry data from several organizations between July and October, they report frequent gaps such as error spans lacking status messages or exception events, and missing or default service names. They also found that many traces consisted of a single span, often signaling broken context propagation. The authors caution that their sample reflects teams that sought telemetry quality help and is not an industry-wide survey.