Explore what's new in AI

OpenAI Codex v0.142.4 adds GPT-5.6 Sol, Terra, and Luna to model picker

Codex users can now switch among Sol, Terra, and Luna from the CLI, which makes the new frontier lineup easy to compare.

Cursor internal screenshot shows GPT-5.6 Sol preview with 1M context

The internal-only preview calls out 1M context and agentic coding, biology, and cybersecurity, which hints at where Cursor is heading.

Gemini Enterprise adds an Inbox for review, in progress, and done work

The sidebar now groups agent output by status, which fits the way teams already triage and close work.

Candidly reads support chats live and steers resolution in real time

Instead of grading after the fact, the model watches the dialogue in progress and suggests actions that move toward resolution.

JetBrains Air brings Codex, Claude, Gemini, and Junie to Windows

Air now runs on Windows and connects several agent backends from one entry point, which makes it easier to swap tools.

Pauliusztin's raw, wiki, outputs folders bootstrap a second brain

The three-folder layout keeps messy source material separate from cleaned knowledge and final work, which makes the system easier to maintain.

H company opens its computer-use agent API in beta

Developers can build on a managed cloud browser and screen-clicking agent model, then expand with Python and TS SDKs.

LangChain adds dynamic subagents that write orchestration code

The main agent can now spawn subagents programmatically, which fits large fan-out jobs over documents and data.

LlamaParse's Retrieval Harness gives agents hybrid search and file tools

It combines vector search, keyword search, grep, and file reads so agents can navigate very large corpora without custom glue.

LangChain shows how to turn a LangGraph agent into a voice agent

The tutorial wires LangGraph into Pipecat and traces the full call path in LangSmith, making voice integration easier to reproduce.

Dexhorthy says to layer small agent loops onto existing systems

The advice is to ship value early by adding review, implementation, and feedback loops one by one instead of rebuilding everything.

Pauliusztin uses an LLM wiki to design a coding harness

The workflow ingests harness repos into one shared wiki, then uses that context to brainstorm, compare, and build faster.

Warp adds custom routers for task-based model selection

Warp lets teams route by complexity or plain English rules while still using BYOK or custom inference URLs.

OpenHands shows a sidekick pattern that cuts LLM spend

A cheap sidekick can handle exploration and routine edits while the main agent keeps expensive tokens for the hard parts.

NVIDIA's HORIZON turns hardware design into repository-level agent work

HORIZON wraps hardware design in executable harnesses and reports full benchmark completion across several suites.

Hamel Husain says hard-to-eval AI products are usually a design smell

If users cannot verify the output, the product design is probably the bottleneck before evals even start.

Zenith harness pushes open models above frontier baselines on hard tasks

Zenith says a multi-model harness can push open models above frontier baselines on long, hard tasks, not just short benchmarks.

IRIS packages Gemini Live plus Hermes Agent into a public repo

The repo turns a voice-driven agent loop into something you can inspect, fork, and improve instead of treating it as a demo.

Google's Paper Assistant Tool automates scientific review

Paper Assistant reads whole manuscripts, checks claims and proofs, and found more math errors than a single model call.

Hermes Agent moves session DB writes off the event loop

The new DB seam offloads blocking writes to worker threads, which keeps the loop and heartbeat responsive under load.

ClinePass bundles open-weight coding models behind one monthly plan

ClinePass wraps curated open models behind a single subscription, so teams do not have to juggle separate API keys.

Cursor launches an iOS app for cloud agents and remote control

Cursor's iOS app lets you launch cloud agents and remote-control local ones, bringing review and diff handling into the phone workflow.

Google AI Studio adds full chat history when sharing apps

The share toggle preserves full chat history when you publish an app, which makes prompts and workflow context far easier to reproduce.

Base44 launches Base 1 as its first web-app generation model

Base 1 is the first model from a vibe-coding platform that now owns the full stack, from backend to model weights.

VS Code and TypeScript team up to ship TypeScript 7

Microsoft and the TypeScript team are treating TS 7 as a dev-loop upgrade, with faster builds and less toolchain friction.

T3 Code mobile lets one app steer work across multiple computers

T3 Code is trying to make one mobile app control work across several machines, which is a nice fit for agents and cloned repos.

ComfyUI's Comfy MCP exposes saved workflows to agents

Agents can run saved ComfyUI workflows, search models and templates, and replay jobs on new inputs without editing nodes.

OpenCode v2 adds a shell manager for backgrounded subagents

The new UI makes it easier to see what is live, background it, or kill it. That is the kind of control multi-agent work usually needs.

Spotify says 73% of PRs are AI-assisted in a 20M-line monorepo

Spotify says 73 percent of PRs are AI-assisted and its judged codemod flow raised success from about 25 percent to 80 percent.

OpenAI teases a Codex shortcuts upgrade for July 15

OpenAI is teasing a shortcuts refresh for Codex. The only concrete date is July 15, but it signals another workflow pass on the CLI.

OpenClaw ships native iOS and Android apps for agent control

Native mobile apps make OpenClaw usable outside the desk, with channels, tasks, and replies exposed on iOS and Android.