Explore what's new in AI

Think orchestrates Claude Code in Cloudflare sandboxes with zero API keys

Think runs Claude Code in Cloudflare sandboxes, one task per container, and routes egress through AI Gateway. The zero-key setup is a clean isolation pattern.

Mark Chen says OpenAI models are nearing self-sustained research

OpenAI's research chief is talking about models doing more of their own innovation and research loops. That's a signal that the next step is less assistant, more autonomous R&D.

Instructor 1.15.4 ships a maintainer sweep of parsing and doc fixes

Version 1.15.4 focuses on small correctness fixes, from streamed JSON backticks to Image.autodetect on raw bytes. Maintainer sweeps like this reduce weird edge cases.

Paper adds selected text layers as prompt presets for image generation

Selecting text layers as prompt presets makes image generation much more reproducible. It lets you save style, mood, and color ingredients inside the canvas itself.

ChatGPT's plugin menu exposes Agent mode, Netlify, Supabase, and Slack

The plugin list shows ChatGPT leaning into a connector hub, with agent mode alongside Netlify, Supabase, Slack, and Monday.com. That's a concrete app-platform signal.

OpenAI adds GPT-5.6 to Codex and Amazon Bedrock

Putting GPT-5.6 into both Codex and Bedrock suggests OpenAI wants developer and enterprise distribution at once. That makes the launch about use, not just announcement.

Towards an Agent-First Web adds identity, purpose, and agents.txt to requests

The proposal gives agents identity, purpose, and policy metadata on each request, plus an agents.txt control file. That could become the web's bot access layer.

Accenture's Memex RL stores long-horizon agent work in indexed memory

Memex keeps long-horizon work in an indexed store instead of one giant context window. That is exactly the sort of architecture agent teams need.

BrowserBC turns browser sessions into reusable skills and lifts web scores

BrowserBC treats browser runs as reusable skills, then reports much better web task success with fewer steps. That's a strong pattern for flaky computer-use flows.

PlanBench-XL tests agents across 1,665 tools and hidden dependencies

The benchmark says agents can still fail hard when tools are hidden behind a huge search space. Tool ecosystems at this scale need planning, not just better models.

Codex desktop shows all remote sessions on mobile and other devices

Seeing all remote sessions from mobile turns Codex into a real command center for in-flight work. That makes handoff and supervision much easier away from the laptop.

DAIR AI's agent reading list packs 11 core papers into one bookmark

Eleven papers in one place makes the agent stack easier to catch up on, from memory and planning to safety and RL. Good bookmark fodder for builders.

Claude Code's /skill invocation behavior keeps changing under the hood

The behavior changing across weeks suggests skill invocation semantics are still fluid. That's a useful warning for anyone building on Claude Code skills.

IRIS pairs Gemini Live with Hermes Agent for hands-free task execution

Gemini handles the conversation while Hermes executes tasks in the background, then hands results back for review. The gesture layer makes the whole loop feel hands-free.

OpenCode signs up for Telnyx by following agent-specific onboarding steps

The bot signup flow is actually optimized for the agent, with explicit instructions and bot checks. That's a strong sign more SaaS onboarding will need machine-readable paths.

OpenAI tests a Codex effort slider and reworks real-time voice support

The UI suggests OpenAI is exposing effort as a slider, not just a hidden parameter. If it ships, teams get a clearer way to trade speed, cost, and quality.

Supervision passes 45K GitHub stars after a 5K surge in three weeks

A 5K-star jump in three weeks usually means the repo is becoming part of the default agent workflow. People building coding-agent supervision are paying attention.