This video shows the closing remarks on the final day of AI Engineer NYC 2026, given by Ksenia Se of Turing Post and swyx. It covers topics from building agents (AI programs that carry out tasks on behalf of people) to how financial institutions are changing the way they work.
This video covers how to use the KV cache (a space where a model stores calculation results during a conversation) as a security signal for AI agents. The speaker is Ari Azarafrooz, founder of Intrinsec AI. The video was posted on the AI Engineer channel on November 9, 2026.
Stanford researcher Maty Bohacek examines how AI agents (AI programs that act on their own toward goals) handle procedures that verify someone is human through phone or video calls. The scenarios include opening accounts, approving money transfers, and recovering access.
Mislav Ivanda of Daytona shares what he learned from adding sandboxes (isolated execution environments) to major agent frameworks. Frameworks that disagree on almost everything have reached the same conclusion.
Jason Mancuso of Modal presents a video arguing that teams running agents already have what they need to start reinforcement learning (RL, a training method that improves models through rewards). It also argues that treating models as swappable parts has limits.
Warp founder and former Google Docs engineering lead Zach Lloyd describes a shift away from developers directing AI agents (programs that carry out tasks on their own) from their laptops, toward isolated cloud software factories. The video demonstrates a tool called Warp Factories and shows how to watch agent runs.
When you hand off work made with an AI agent (an AI program that does tasks on a person's behalf) to a teammate, important context often gets lost. This talk covers what goes missing and shows one way to fix it using a tool called Delta. The video was posted on the AI Engineer channel.
This talk covers how AI agents copy accessibility problems found in the code they learned from. Accessibility means making products usable for people with disabilities without barriers. In an experiment, the speaker says an agent given accessibility rules finished the same task faster and at lower cost.
Ashwin Paranjape of Samaya AI explains why forecasting company earnings is a better test for AI finance agents than predicting stock prices. He says the test ran on more than 450 earnings announcements in an environment that blocked access to future data.
Tom Moor of Linear says writing code is no longer the bottleneck. He argues that as agents (AI programs that work on tasks for people) build more products, keeping products consistent becomes harder.
Google DeepMind says it used its AI agent framework AlphaProof Code to formally verify the core security properties of the Signal messaging protocol at the machine-code level. The announcement was made on November 10, 2026.
Google Antigravity, its AI work tool, now has a new screen mode that is easier for non-developers to use. Users can switch between it and the original developer screen. Both use the same AI agent, which can handle multi-step tasks on its own, so the features stay the same.
AI agent and hardware startup Malabs has been selected for TIPS, a tech startup support program run by the Ministry of SMEs and Startups. It plans to use 800 million won in R&D support to grow its AI technology and expand its business. The company secured both investment and government support about seven months after being founded.
Meta is steadily expanding its AI agent Muse with real-time voice conversation, an avatar that copies facial expressions and gestures, and Spaces, a workspace organized by topic. However, the rollout of Meta's AI terms has been indefinitely postponed over controversy about collecting news without permission, and concerns about personal data exposure have also been raised.
Cursor's voice chat feature lets you talk with a coding agent (an AI that makes code changes for you) as if you were on a phone call. After giving a task, you can ask during the call what has changed and whether it is still in progress. This was announced in the November 6 changelog.
OpenAI has released a report evaluating the ability of AI agents to work in wet labs (laboratories where people directly handle liquids and reagents). The report says the AI showed a certain level of ability to create experimental procedures and find errors in its steps. Both expectations and concerns are growing.
As AI agents (artificial intelligence programs that handle tasks on their own) become more common, technology to manage them safely is becoming important. Nvidia's blog introduced how Korean companies are releasing security tools and AI models to keep up with this trend.
Anthropic published how its 'dreaming' technique works and the results of its experiments. In this technique, AI agents review past work records during idle time and correct false memories. The company found that keeping only important information worked better than storing everything.
This talk covers a case of applying an AI agent (a program that carries out multi-step tasks on its own) to indexing work for biopharma documents. The speaker is Matthew Jewkes, who built the system at Clarivate. The early approach reached only about 80 to 83% accuracy.
DeepSeek's AI agent platform, DeepSeek Harness, became a service used by 1.54 million people a day just three months after its launch. External developers have built about 35,000 plugins for it. Based on this growth, DeepSeek plans to launch an official marketplace and expand support for developers.