Explore what's new in AI

TinyFish BigSet turns a sentence into a verified dataset

This is a useful agent pattern for data work: describe the dataset, let agents research and dedupe sources, then export a clean table.

Harness-1 moves working memory out of the policy

Search agents get better when the harness stores candidate pools and evidence instead of making the policy track everything.

Codex Mobile adds Face ID, SSH to Windows, and side chats

Codex Mobile is moving toward a real control surface, with better auth, side conversations, and remote Windows access for follow-up work.

OpenAI and Mayo Clinic co-train a frontier model for healthcare

The partnership shows big labs are moving deeper into regulated domains. Healthcare is becoming a real test case for frontier model deployment.

GEPA can optimize any agent CLI by auto-tuning prompts

Wrap an agent in a function, let GEPA tune the prompt, and you get a cheap way to improve behavior without rewriting the whole stack.

Browser Use can start 25 browsers in under a second

Fast browser fleet startup matters for agentic testing and parallel web work. This is a practical benchmark, not just a demo.

AI still forecasts scientific progress poorly, despite strong domain knowledge

The models recognize plausible research paths better than they predict real breakthroughs. Hindsight is still much easier than foresight.

AutoMedBench says frontier models still fail medical research workflows

Models can set up pipelines, but validation and engineering still break the chain. Medical auto-research is still far from reliable.

Microsoft Web IQ rearchitects search for AI agents

Microsoft is treating agent search as a distinct product category, with grounding APIs tuned for quality, latency, and token efficiency.

OpenCode wins official OpenAI support after Anthropic blocked it

A platform ban became leverage for a faster ecosystem deal. The bigger lesson is that agent tools are racing toward subscription interoperability.

Claude Code workflows let one prompt orchestrate subagents and QA

Workflows are turning Claude Code from an interactive shell into a small orchestration layer, with verification built into the loop.

Claude Code adds a self-check loop before handoff

Encoding manual review into the workflow lets Claude catch its own mistakes before you see them, which is the right direction for agents.

Codex can recreate animations from a video prompt

This is a concrete multimodal workflow, feed Codex a video and ask it to rebuild the motion in another tool. Very practical for motion design.

Stanford's Pupper robot dog runs Gemini Live and Robotics-ER

A real robot demo with open hardware is more useful than a flashy clip. It shows how multimodal models can drive on-device perception and reaction.

NVIDIA unveils an open humanoid robot platform with Isaac GR00T and Unitree H2

Physical AI is moving into open robot stacks, not just demo videos. Open hardware plus open models could speed up iteration a lot.

Crafter and CraftEditor turn figures into editable, agent-friendly outputs

Scientific figures become structured artifacts instead of dead images, which makes targeted edits and reproducibility much easier.

TanStack Start adds Rsbuild 2 support alongside Vite

Start now reaches deeper into the Rspack and Webpack ecosystem, which makes the framework easier to adopt in existing build pipelines.

Warp adds SSH file tree and code review support for remote sessions

Remote agent sessions now feel much closer to local dev, with file editing and code review over SSH. That reduces the gap between cloud and laptop workflows.

Uncle Bob Martin describes agent-assisted QA over UI and gherkin

Agents can now write QA scripts, execute UI tests, and even generate repeatable test harnesses. That's a concrete shift in how teams can maintain UI coverage.

Manus adds native MCP connectors and Skills to its chief-of-staff agent

Native connectors make a chief-of-staff agent much more practical, because it can actually reach the tools and data your team uses.

NotebookLM adds Planning Mode for Video Overviews

NotebookLM is moving from passive summaries to guided planning for video. That's a useful sign that document and video workflows are converging.

Google DeepMind opens Co-Scientist to individual researchers

Co-Scientist is no longer locked to select partners, giving solo researchers a multi-agent hypothesis loop for science work.

OpenAI plans to merge ChatGPT, Codex, and Atlas into one desktop app

OpenAI is recasting Codex as a broader work surface. If this lands, ChatGPT becomes the default shell for more knowledge work.

Perplexity Search as Code lets agents compose search workflows

Writing search flows in Python lets agents parallelize queries, dedupe results, and rank before context hits the model. That's a useful abstraction.

Centaur makes web search and extract free by default

Removing the API-key barrier makes it much easier to try search and extraction in agent flows, especially for quick prototypes.

Cursor publishes durable-execution lessons for cloud agents

Cursor is saying the quiet part out loud, cloud agents need durable execution, a harness, and realistic dev environments, not just a server.

LangChain Studio adds one-click deploy for agents

This shortens the jump from prototype to production. For agent teams, deployment friction is now part of the product.