Prasanna Sankar Says Cloud AI Agents Handle 20–30 Tasks

Key Takeaways

• New York is creating an AI safety research center at SUNY Binghamton, with about 80 researchers and Cornell’s Greg Morrisett as co-lead; Gov. Kathy Hochul says she will explore stronger state rules.
• Ohio’s eight AI bills are unlikely to pass before the mid-December deadline; one proposal would penalize tools that encourage self-harm or violence.
• In a UK survey of 2,000 adults, 84% wanted limited or no AI-agent control; 36% worried about data theft and 28% about unauthorized spending.
• LinkedIn’s career agent still requires human review of prompt changes; its automated evaluator agrees with reviewers 55% of the time at the lowest score and 80% at the highest.
• YC founder Prasanna Sankar says cloud coding agents can run 20–30 parallel tasks versus 3–5 for desktop coding, but warns of pull-request overload, alerts and token costs without clear oversight.
• Jemima Abu recommends semantic HTML, keyboard access, design systems and automated accessibility audits for agent-generated code; she says her Figma test was not empirical proof.
• German firms Herding and Safelog are using accumulated production data and an internal AI platform, while other manufacturers face costs, red tape and data-readiness gaps.
• Elkhart firefighters use AI-and-sonar AquaEye to scan up to a football field long and more than 100 feet deep, helping guide divers to possible victims or objects.
• Marktechpost’s free GPU Leaderboard and Chrome extension compare cloud GPU on-demand and spot prices by GPU hour.
• A study analyzing data from more than 100,000 patients found AI-derived inflammation scores may identify cardiovascular risk even when coronary calcium is low or absent.

New York launches independent AI safety research center

Gov. Kathy Hochul is creating the New York Center for AI Responsibility and Research at SUNY Binghamton. About 80 researchers will independently test AI systems for safety and security, including whether they can be manipulated, produce dangerous results, or threaten privacy. Cristin Frodella of Common Sense Media said AI can pose risks to children, while Cornell Tech’s Greg Morrisett, a co-chair, said the center will study ways to prevent failures and attacks. Hochul said she plans to explore stronger AI safety laws in New York in the coming months.

Greg Morrisett will co-lead New York AI safety center

SUNY has established a statewide artificial intelligence safety research center launched by Gov. Kathy Hochul, with computer scientist Greg Morrisett serving as co-lead. Morrisett, a professor of computer science at Cornell University, will work with other experts to develop AI systems that are safe and trustworthy. The center will study AI safety and security, with possible effects on fields such as healthcare, finance, and transportation. Its goal is to help build a safer and more trustworthy AI ecosystem.

Elkhart firefighters use AI device to find people underwater

The Elkhart Fire Department is using AquaEye, a device that uses AI and sonar to help locate possible victims underwater. It can scan up to a football field in length and more than 100 feet deep, and marks possible people on a map while showing submerged objects differently. Fire Division Chief of Operations Jason Gour said the device can reduce searches from hours to minutes. Crews use its location information to guide rescue divers to a person or object in the water.

Coding agents shift the bottleneck to reviewing and merging work

Prasanna Sankar, a three-time YC founder, described a system that lets cloud coding agents work in copied laptop environments on raw EC2 machines, with ports, secrets, authentication, and test data set up. Agents plan and split tasks, run browser and end-to-end tests, and provide recordings and staging links for review. Sankar said cloud agents can handle 20 to 30 parallel tasks, compared with 3 to 5 for desktop coding, though his own context broke down beyond 10. He also warned that poorly managed agents can create too many pull requests, constant alerts, and high token costs, so he recommends owners, budgets, check-ins, and visible work for each bot.

Accessibility rules can help coding agents work faster and cost less

Jemima Abu said accessibility requirements can help agents produce better code, while noting that 96% of the top 1 million websites have accessibility issues and that 16% of people worldwide live with a disability. She recommends starting with semantic HTML, keyboard access for interactive elements, and native features such as details for accordions and popover for dialogs. Abu also advises teams to use a design system, review their codebase, define requirements early, and automate audits so agents do not repeat bad patterns. She said her Figma test was a one-off, not empirical proof, and warned that inaccessible code can drive away customers and invite lawsuits under laws such as the Americans with Disabilities Act.

German manufacturers turn decades of data into an AI advantage

German manufacturers are using years of production data to improve AI tools and stay competitive. Herding, an industrial filtration company with about 400 employees, has collected technical details from thousands of projects since it began making filtration units in 1977. Managing director Fabian Schünke says this data could help the company develop new solutions. Robot maker Safelog has organized its work around an internal AI platform that helps staff find production problems and develop products. Many German businesses still face high costs, red tape and limited digital tools, and some have not collected data in a form AI can use.

Ohio AI safety bills face a mid December deadline

Ohio lawmakers are unlikely to pass AI rules before their mid December deadline, even as lawmakers from both parties push for action. A bill from state Reps. Christine Cockley and Ty Mathews would penalize developers whose AI tools encourage self harm or violence. Eight AI bills are before the House Technology and Innovation Committee, but chair Thad Claggett says none is ready for a vote because of possible problems. Cockley says pressure from technology companies may be slowing action, while she hopes new leadership under Governor elect Amy Acton will help next year. Gov. Mike DeWine has repeatedly called for safety rules.

UK consumers want control over AI agents, survey finds

A survey of 2,000 UK adults found that 84% would give AI agents only moderate or low control over their lives, or none at all. The research, conducted by Censuswide for Eskenzi PR from September 16 to 18, 2026, found that 36% worry their data could be stolen and 28% fear an agent could spend their money without permission. Many people prefer AI to prepare tasks for approval or handle routine work while asking permission for important decisions. Forty percent said the AI developer should take primary responsibility if an agent is hacked or acts without permission. The research says clear permissions, secure sign-in, monitoring and human oversight can help build trust.

Marktechpost launches cloud GPU price tracker and Chrome extension

Marktechpost has launched GPU Leaderboard, a free tool that compares cloud GPU prices for AI inference. It ranks rentable GPUs by price per GPU hour and gathers on demand and spot prices from providers’ own pricing pages. The company also released GPU Leaderboard: Deploy It, a Chrome extension that provides the same pricing data. The tool aims to help users find which GPU they need and where to rent it for less.

LinkedIn’s self improving AI agent still relies on human review

LinkedIn says its career agent can improve its prompts through a seven step process, but people still review changes before they reach production. The agent uses a member’s ID, search history, profile and preferences to suggest details such as preferred job titles. LinkedIn’s evaluation judge agreed with human reviewers about 55% of the time for the lowest score and 80% for the highest, and had trouble judging some areas such as domain expertise. The team uses reviewer disagreements to improve its scoring rules, and it runs the process weekly or every two weeks. Broader changes to the workflow and code remain future steps.

AI inflammation scores reveal heart risk beyond calcium scans

A study published in the Journal of the American College of Cardiology found that AI derived inflammation scores can help predict cardiovascular risk, even when coronary calcium levels are low or absent. Researchers used a machine learning algorithm to analyze data from more than 100,000 patients. The findings suggest these scores could help identify people at elevated risk who may not show obvious signs of disease. The researchers hope the study leads to more research on using AI scores to diagnose and treat cardiovascular disease.

Sources

• Gov. Hochul establishes New York Center for AI Responsibility and Research
• Greg Morrisett to co-lead new NYS AI safety research center
• Elkhart Fire Department uses new device with AI to help locate victims in water
• Merge Is the Bottleneck for Coding Agents
• Accessible Agents Cost Less and Ship Faster
• ‘Treasure trove’ of manufacturing data to power German AI
• Ohio likely won’t pass AI regulations before year ends
• AI Agents Promise Convenience, But Are We Giving Them Too Much Control?
• Marktechpost AI Launches GPU Leaderboard and the Deploy It Chrome Extension
• LinkedIn's Self-Improving Agent Still Needs Humans
• AI-derived inflammation score improves cardiovascular risk prediction beyond CAC