Explore what's new in AI

Cursor 3 adds Design Mode for targeting browser UI elements

The new mode lets you annotate browser elements directly, which is cleaner than pixel hunting for brittle UI automation.

Summarize 0.13 adds local video slides and better GPT-5.4 routing

It can now summarize local slide videos, and the new model routing makes media and streaming inputs less fiddly for CLI users.

Vercel Emulate adds email inboxes for magic link and OTP testing

You can test signup, magic-link, and OTP flows without sending real mail. That removes one of the slowest parts of auth QA.

Portless adds --lan so phones on Wi-Fi can hit local dev apps

The new LAN mode exposes local apps on your network with HTTPS intact, which makes real-device testing much less painful.

Unsloth publishes Gemma 4 fine-tuning notebooks for 8GB VRAM setups

The notebooks make Gemma 4 tuning feasible on small hardware, with free Colab paths and lower VRAM requirements than stock training.

Claude Code 2.1.94 raises default effort and fixes hook and login bugs

The release fixes rate-limit hangs, broken macOS login, and ignored plugin hooks while nudging more users to higher-effort runs.

Exa adds x402 support so agents can pay for web search in USDC

Agents can now fetch search results with a 402 flow and settle in USDC, removing API keys from many retrieval workflows.

Cognition ships SWE-1.6 with more parallel tool calls and fewer loops

Cognition says SWE-1.6 calls tools in parallel far more often, which should mean fewer reasoning loops and faster repo work.

Superset pitches a parallel-agent IDE for running coding agents together

Superset wants to be the IDE for running coding agents side by side, so you can keep multiple tasks moving instead of waiting on one model.

Zed's Zeta2 gains a 30% better acceptance rate with a better teacher

Zeta2 improves code completion acceptance by 30% without just scaling the model. Zed says a better teacher and better traces did it.

Hugging Face publishes Hermes Agent reasoning traces for tool-use training

The dataset gives agent trainers real multi-turn tool-use trajectories, not synthetic prompts. That is useful fuel for better harnesses.

Stanford shows single-agent LLMs beat multi-agent systems at equal budget

When thinking tokens are matched, single agents win more often. A lot of multi-agent gains look like hidden compute, not architecture.

ComfyUI publishes a node workflow to recreate the viral inflation effect

ComfyUI turned a viral image effect into a reusable node graph, so others can recreate the look instead of reverse-engineering it from scratch.

Google AI Overviews show why answer quality is hard to measure

The same answer can be wrong, right for the wrong reason, or hard to source back. That makes search-quality evals much messier.

Box Agent fills out an RFP response by searching the knowledge base

The demo shows Box Agent extracting questions, pulling sources, and generating a full Word doc while the user does something else.

DeepMind's AlphaGenome traces disease mutations in non-coding DNA

The model targets the hardest part of genetics, predicting which variants matter before wet-lab work. That makes it a real science model, not a demo.

Claude Code skill adds source-cited deep research over complex documents

The skill turns PDFs, Word docs, and slides into an auditable report with word-level citations and bounding boxes back to source.

Google ships AI Edge Eloquent, an offline dictation app for iPhone

The app auto-polishes dictated text and supports custom dictionaries, which is a surprisingly useful fix for a long-standing mobile UX gap.

Browser Use turns CAPTCHA into a gate that only agents can solve

If an agent clears the challenge, it gets free API access. That is a clever pattern for bot-gated services and internal demos.

Claude Code prompt repo reconstructs hidden system and tool prompts

The repo surfaces the hidden prompts behind Claude Code, making its planner, memory, and verifier behavior much easier to study.

Warp adds rendered Markdown tables to its terminal editor

Rendered tables make Markdown less annoying in-terminal, which matters if you inspect docs and data without leaving Warp.

AI Battle turns Claude vs Codex into a judged problem posing competition

The flow asks two agents to invent hard questions for each other, then scores the answers. That's a clever stress test for creativity and verification.

Render Workflows turns one annotation into async orchestration

One annotation now covers retries, parallelism, nesting, and observability, simplifying durable agent pipelines and background jobs.

Browserbase argues agents need a full web stack, not just APIs

The platform pitches one API for browsers, search, fetch, sandboxes, and models, which is a more realistic starting point for web agents.

Architect turns one prompt into a multi-agent workflow with UI and scheduler

The demo shows an orchestrator spawning subagents, selecting model endpoints, and wiring a manual trigger UI plus a scheduler.

S3 files let agent swarms share one filesystem without a sandbox VM

The argument is that POSIX tools plus S3 can replace a VM for many agent swarms, letting you fan out far more compute on the same storage.

Claude Code's /autofix-pr command starts autofix from the terminal

A single CLI command now kicks off PR autofix after you finish a change. That cuts one more bit of review busywork from the loop.

OpenCode adds Search and Extract plugins for browsing and parsing

The plugin split gives OpenCode a cleaner search and fetch path, so agents can get web context with less custom glue code.

OpenAI says Codex now blocks cyber abuse by passing user IDs downstream

The team says it can reject bad actors more reliably when user IDs are attached to requests. That's a concrete abuse-prevention pattern for agent platforms.

Ideogram Layerize adds editable text layers and transparent PNG output

The API now turns flat outputs into editable layers, which makes localization, compositing, and design reuse much easier.

VS Code Learn launches a guided course for agent-first development

Microsoft is shipping a guided course that teaches agent-first app building inside VS Code, which lowers the barrier for new builders.

CodexBar 0.20 adds provider switching and cost history merges

The menubar tracker now supports more providers, lets you swap Codex accounts without relogin, and de-dupes session costs.

MiroMind turns market analysis into a cited workflow over news and charts

The thread shows a repeatable workflow for market research, with sources, charts, and forecast framing instead of a black-box answer.