Explore what's new in AI

Gemini iOS redesign surfaces Canvas, Deep Research, and media tools

Google is pushing a broader mobile surface, with creation, research, and media modes exposed in the new layout. That suggests Gemini is becoming a hub, not just a chat box.

David Soria Parra says MCP servers should be products, not REST wrappers

The thread argues MCP only pays off when servers are built around user journeys, not raw API wrappers. Useful semantics beat generic endpoints.

OpenAI GPT-5.5 makes Gmail search in ChatGPT faster and more accurate

The practical win here is boring but real, better Gmail search. It shows where GPT-5.5’s value lands for everyday agentic work.

Notion adds Discord and Google Drive triggers for custom AI agents

Notion is widening agent entry points beyond the app itself, so AI can wake up from Discord commands or file changes in Drive.

CopilotKit maps generative UI from controlled flows to open runtime

The library exposes the control tradeoff directly, with shared state and frontend tool calls as the core primitives for agentic UI.

ClickHouse can store LLM traces, but regex search is the bottleneck

The thread argues ClickHouse is fine for traces at scale, and that the real pain point is search ergonomics rather than raw storage capacity.

dbreunig sketches six commandments for coding with agents

The list is a compact checklist for sane agent workflows: implement to learn, rebuild often, keep intent documented, and verify end to end.

Rosie adds versioning and lockfiles for agent skills

A skills package manager with lockfile support makes agent setup more reproducible and gives teams a way to pin behavior instead of hoping.

DeepAgents virtual filesystem stages context without a full sandbox

A virtual filesystem gives agent builders a way to stage files and context-management hacks without spinning up a full execution environment.

Contextual Agentic Memory says vector stores are memos, not memory

The paper argues retrieval-only memory cannot consolidate or generalize, which leaves long-running agents brittle and more exposed to poisoning.

OpenAI Symphony routes Linear issues into autonomous Codex work queues

OpenAI says some teams saw a 500 percent PR increase by polling issue trackers, isolating agent workspaces, and auto-opening follow-up tasks.

Magnific workflow chains three models into a repeatable video recipe

The workflow uses a reference image, a motion sheet, and Seedance 2.0 to generate consistent AI video. That makes it a concrete multistage recipe, not just a flashy demo.

Google AI Studio adds point-and-click image editing to Build

The new edit tool moves image tweaks out of chat and into direct selection, making iterative multimodal edits faster for builders.

Google's DolphinGemma article explains LLMs for dolphin communication

Google is pushing multimodal models into scientific decoding, where the goal is animal communication rather than chat or code generation.

Skeptrune's Claude Code mobile guide wires Termux and Tailscale together

The guide shows a practical phone-first setup for Claude Code using Termux and Tailscale, so you can keep agents running away from your laptop.

Cursor Multitask mode runs several agent jobs in parallel in one chat

The feature lets users queue prompts and have subagents work at once, which is a real productivity bump for long debugging or database-heavy sessions.

Patrick Debois says context should be engineered like code

Debois proposes a lifecycle for context, Generate, Evaluate, Distribute, Observe, giving teams a concrete way to version prompts and test them.

Wiki Builder turns Claude Code into a reusable research wiki generator

The plugin scaffolds a research wiki from raw notes, then keeps the index and linked pages organized around a per-wiki config file.

Pi adds boomerang mode to rewind prompts while keeping file changes

Boomerang mode runs a prompt, then rewinds the session with edits intact and a summary in feed. That makes long agent loops feel less lossy.

RepoBar 0.4.0 caches GitHub data locally and shows rate limits

Persistent SQLite caching trims wasted API calls and makes repo state feel instant, while a visible rate-limit view helps explain stalls.

LlamaIndex says PDF parsing is still a hard problem for agents

The talk explains why PDFs stay messy for document AI, and why VLM-based parsing still matters when agents need reliable text and table extraction.

Kevin Murphy's RL overview bridges classic RL with LLM training

Murphy’s survey is a strong modern RL reference because it connects classic algorithms to RLHF, world models, and multi-turn LLM agents.

BIGMAS improves reasoning with specialist agents in a shared workspace

The paper argues that better system structure can outperform a single long chain of thought. It reports gains from 12% to 30% on some tasks and 57% to 93% on Tower of London.

OpenAI switch to Codex imports settings, plugins, and project config

The migration path lowers setup friction by moving an existing coding environment into Codex with a few clicks. That is the kind of UX that helps agents spread.

Google tests an Omni video model inside Gemini before I/O

Gemini appears to be getting video-generation plumbing alongside the new UI. If this ships, it broadens Google’s consumer model stack beyond chat and images.

Google Flow beta brings image and video generation to Android

Android testers can now use Flow for both image and video generation. That points to a native mobile workflow for creative model use, not just web access.

Google AI Studio Build adds GitHub and Figma imports for app scaffolding

Importing GitHub or Figma context directly into Build cuts setup work and makes AI Studio feel closer to a real app starter than a demo toy.