Explore what's new in AI

LiteParse claims a faster PDF parser with bounding boxes for audits

Bounding boxes make parsed text auditable, which is exactly what coding agents need when they have to cite the source page.

VS Code's Agents window orchestrates sessions across projects

The new window gives you one place to manage agent sessions across projects, with worktree isolation baked in.

TanStack AI adds client and server context to tool middleware

The update makes useChat flows more typesafe by threading context through tools and middleware on both client and server.

Google releases Gemma skills for agent builders

The new skills catalog packages model-specific guidance and setup into reusable assets for Gemma-powered apps.

Codex can now watch Slack, deploy on time, and clean itself up

Codex automations are reaching ops territory, handling weekend checks, timed deploys, and cleanup without handholding.

DesignMD turns any public URL into a DESIGN.md for coding agents

DesignMD extracts live tokens, breakpoints, and interaction states so agents get a real spec instead of a screenshot guess.

OpenAI Voice Hack Night crowns a voice-first agentic OS for phones

The winning demo points toward a mobile future where users talk to a phone OS that can answer and take actions directly.

Blackstone adopts TextQL to expand analytics across portfolio companies

Blackstone's tech lead says the product could help portfolio companies, a good enterprise signal for an analytics tool.

HappyRobot says FDE-style work helped it win major freight brokers

The case study shows enterprise AI is often won through deployment work, not just model quality, especially in high-touch verticals.

LangChain's sandbox keynote shows how serious agents should run code

The talk lays out the runtime agents need, isolation, network controls, and checkpoint restore when things go wrong.

turbopuffer's tokenizer hits 423 MiB/s with a DFA and FWHT rotation

The tokenizer rewrite shows how much CPU search systems can save once tokenization stops being treated as a commodity step.

LobeHub's Chief Agent Operator hires and schedules agents 24/7

The pitch is an operator layer, not another chatbot, with cloud scheduling and Slack or Discord reporting back to the team.

Perplexity explains why vector indexes are components, not full databases

The guide draws a clean line between search algorithms and production vector databases, which helps teams avoid scaling traps.

Cursor's developer habits report says AI doubles speed and widens the power-user gap

Cursor's report puts numbers on AI coding, showing faster cycles, bigger PRs, and a widening advantage for power users.

Andrew Ng says AI FDEs are rising, but AI engineers still matter more

Ng says embedded FDEs are growing fast, but the bigger demand is still for in-house AI engineers who can own workflows and stay vendor-flexible.

Google DeepMind shows how to prompt Gemini Omni for video edits

The guide says Omni works best with intent, framing, lighting, and iterative edits, not frame-by-frame micromanagement.

React Native Doctor scans apps for bugs and fixes them with agents

The CLI finds perf and security issues in React Native apps, then uses agents to patch them and shorten the review loop.

OpenCode says it hit 10M monthly actives and 5T tokens a day

OpenCode's May numbers make the scale hard to ignore, with usage that looks more like infrastructure than a hobby project.

Google AI Studio Build connects Gmail, Drive, and Sheets inside the app

AI Studio Build now reaches into workspace apps, letting new apps pull live context without sending users across tabs.

Unsloth shows how to pair local LLMs with MCP servers

The guide connects local models to MCP servers so agents can use tools and APIs privately instead of through a cloud model.

Tencent Hunyuan ships Hy-Memory as a second brain for long-lived agents

Hy-Memory tries to fix long-session fragmentation with a compact memory stack that claims lower tokens and faster updates.

ComfyUI adds 11 image, video, audio, and 3D integrations in May

The monthly roundup shows how fast the ecosystem is expanding, with new models landing across image, video, audio, and 3D.

Effective Feedback Compute scores agent harnesses by useful feedback

The paper argues raw tokens miss the point, the right metric is whether feedback was useful enough to change future decisions.

OpenAI retires o3 and GPT-4.5 from ChatGPT

OpenAI is pruning older models from ChatGPT, which keeps the lineup leaner but forces teams to track deprecations closely.

JetBrains Mellum2 lands in vLLM with 128K context and tool calling

The model brings long context and tool use to an open-source MoE, which makes it easier to run agentic code workflows locally.

LangChain demos a macro research agent that produces cited briefings

The example shows Deep Agents doing real analysis, anomaly detection, and source-cited reporting in one workflow.