Explore what's new in AI

Cursor adds multi-repo automations in one agent workspace

Multi-repo support lets agents reason across codebases, test changes, and verify fixes without bouncing between separate windows.

Dwarkesh and Eric Jang map why MCTS still matters for language agents

The discussion shows where explore-exploit credit assignment helps, and why language models still struggle to use MCTS well.

Vtrivedy explains how to use evals as the backbone of agent shipping

The thread turns evals into regression tests, training data, and a product contract, which is a useful mental model for teams.

OpenHuman combines memory trees, app imports, and a Google Meet participant

This repo is more than a chat wrapper, it packages persistent memory, app data import, and an in-meeting agent surface.

Grok Build 0.1 lands in Vercel AI Gateway and OpenCode

xAI's coding model is now showing up in multiple agent surfaces, which makes it easier to test the model outside one vendor UI.

Cloudflare publishes Agents repo and docs for browser-native app builders

The repo and docs give teams a reference for hosting agent apps on Cloudflare instead of wiring everything from scratch.

LiteParse gives due-diligence agents box-level citations for finance docs

Exact bounding boxes and audit trails make document agents far more defensible for financial workflows than plain OCR text.

Vercel ships an AI Gateway plugin for WordPress model calls

WordPress sites can now call text, image, and video models with one API key, which makes AI wiring much less bespoke.

Fireworks Agent can fine-tune a Qwen model from natural language

Pairing Claude Code with Fireworks Agent turns post-training into a prompt-driven workflow, with dataset and skill files in the loop.

OpenAI says ChatGPT Record was built by a 5-person team with Codex

A tiny team shipped a native macOS recording feature using o1, o3, and early Codex, a good signal for AI-assisted app development.

Meta's leaked all-hands says engineer traces will train coding models

The leak suggests Meta wants internal code, clicks, and edit traces to become training data, turning workers into model supervision.

Verbalized sampling restores creativity with a 20-word prompt tweak

Asking for multiple candidate answers and probabilities can recover much of the diversity lost to alignment, without retraining.

LangChain's code interpreter runs code without a full sandbox

A lightweight execution environment is enough for many RLM and tool-calling workflows, without spinning up full infra.

PaddleOCR 3.5 runs on Hugging Face Transformers as a backend

Moving OCR into the standard Transformers stack lowers friction for RAG and document-AI systems that already live there.

ComfyUI adds Stable Audio 3.0 for licensed music and SFX generation

Day-zero support means audio creators can drop open-weight music models directly into existing node workflows.

SID-1 beats RAG plus rerank on recall while running 24x faster

RL-trained search over big corpora is getting cheaper and better, and the 99% cost claim makes this especially notable.

Exa raises $250M to build search infrastructure for agents

Exa says 5,000 companies and 400,000 developers already use it, which makes agent search a real infrastructure category.

Devin adds Slack auto-triage with a three-click setup

Slack auto-triage is a concrete move from chat to operations, especially if you want repetitive channel noise routed into agent workflows.

Warp lets a local plan spawn a cloud team of agents

Warp is pushing coordination up a level, with isolated cloud agents, message passing, and multiple harnesses working on one PR.

Plannotator adds a long-running surface for parallel AI code reviews

Persistent review sessions let teams keep multiple agent critiques open at once, which is useful for large stack-wide changes.

DiffsHub streams million-line GitHub diffs with virtualized CodeView

This is a practical fix for PRs that overwhelm the browser, with streamed rendering and annotations for huge change sets.

Linear Releases tracks deployment state alongside every issue

Bringing version, environment, and status into the issue view gives humans and agents the context they need to reason about releases.

Claude Code 2.1.146 fixes background sessions, MCP paging, and updater bugs

This is a maintenance-heavy release, but the fixes matter for people running Claude Code on Windows, paginating MCP servers, or long background jobs.

Factory's Droid adds Deferred Context Engine for tool-heavy agents

Loading tool schemas only when needed can cut noise and latency for agents that juggle lots of MCP servers and plugins.

MeMo adds a learned memory model behind the LLM

The paper argues memory should be a separate trained subsystem, not a vector-store hack, if you want continual learning without forgetting.

Perplexity cuts context tokens 70% with query-aware compression

The compressor strips ads and navigation before model handoff, so Perplexity can keep answer quality up while feeding fewer tokens.

Hermes Agent hooks into Browserbase's browser skills hub

Reusable browser skills let agents borrow proven site workflows instead of relearning each website from scratch.

Hermes Agent adds skill bundles for one-command capability loading

Bundling skills makes repetitive agent workflows easier to bootstrap and far less brittle when a task always needs the same toolset.

Google DeepMind adds Science Skills to Antigravity for genomics workflows

Google says the science stack pulls from 30 plus life-science sources and sped up a rare-disease analysis. That's real workflow payload, not just a demo.

Google Gemini Omni turns chat into video creation and editing

Gemini Omni makes video generation feel conversational, which matters because it folds text, audio, and images into one editing loop.

ChatGPT app adds shipment trackers, expandable prompts, and a file library

The app is becoming a workflow surface, with shipment tracking, expandable prompts, and a reusable file library all in one place.

Google AI Studio turns prompts into Android apps with one-click install

AI Studio now builds Android apps you can install to a device, and Google says mobile creation is coming soon.