Explore what's new in AI

The Last Human-Written Paper proposes agent-native research artifacts

The paper argues research should be executable, traceable, and grounded in evidence, not just polished prose. That could change how AI papers are reproduced and extended.

OpenAI's Computer Use docs spell out macOS permissions and GUI task limits

The docs make clear this is for supervised GUI work, not free-form autonomy. The permission model matters if you want safer desktop automation.

Pika's Claude plugin adds Podcast, Explainer, and UGC Ad skills

The plugin turns Claude into a persona-driven creative agent with ready-made media workflows. That broadens MCP beyond utility into content production.

Warp adds a /feedback skill that turns bug reports into high-quality GitHub issues

The skill pushes users from complaint to contribution, with an agent drafting the issue and inviting a fix. That is a practical workflow pattern for open-source teams.

Pi v0.72.0 adds Xiaomi MiMo as a first-class provider and changes model metadata

Pi is expanding provider support while breaking old model metadata in the process. That is a sign the multi-provider harness is still evolving fast.

OpenAI tests ChatGPT screen sharing on Android with bubbles

Android screen sharing via bubbles is a lighter-weight agent UX than full casting. If it holds up, it should feel much cheaper to run.

Google Maps adds agentic restaurant booking through AI Mode and Ask Maps

Users can describe vibe, party size, and dietary needs, then get real-time availability and direct booking links. That cuts a lot of reservation search friction.

Chat SDK adds native Teams streaming, Slack Socket Mode, and custom endpoints

Chat SDK is expanding the same bot codebase across multiple chat surfaces. That lowers integration cost for teams that need one agent across Slack, Teams, and more.

OpenAI adds a Custom Dictionary feature for Codex and ChatGPT voice dictation

Adding domain phrases to dictation is a small change with big workflow impact. It removes one of the annoying reasons people still bounce to separate dictation apps.

Recursive Multi-Agent Systems keeps collaboration in latent space

The paper shows latent-space coordination can beat text debate with 8.3% accuracy gains and up to 75.6% fewer tokens. That matters at production scale.

Microsoft Research trains agents on 1,000 synthetic computers for long-horizon tasks

Synthetic computers let Microsoft create month-scale agent traces without waiting on human log data. That could become a major source of training signal for long-horizon agents.

Can AI Agents Agree? says current multi-agent groups still fail simple consensus

The paper shows agreement is still brittle even when every agent is trying to help. That is a real constraint for any workflow that assumes a team will self-coordinate.

PostTrainBench says GPT-5.5 still trails Opus 4.7 inside Claude Code

GPT-5.5 can look strong on one benchmark and still lose inside another harness. That is a good warning against overreading single-number comparisons.

Google asks builders to code the I/O countdown in Gemini Canvas

Google is crowdsourcing a keynote countdown as a creative coding challenge. It is a neat signal that AI Studio is being positioned as a public prototyping lane.

Notion adds Discord and Google Drive triggers for custom AI agents

Notion is moving custom agents from polling to event-driven triggers. Discord commands and Drive changes are a cleaner way to wake agents up at the right time.

VS Code adds Agent Debug Logs for tool calls, model requests, and agent flow

The new logs make agent sessions inspectable instead of mystical. That should save time whenever a chain of tools or prompts goes sideways.

Vercel AI CLI pipes image, video, and text generation through the terminal

The CLI turns multimodal generation into composable shell pipelines, which is exactly the kind of interface agents can automate cleanly.

Anthropic schedules Code with Claude livestream and hands-on sessions next week

Anthropic is using the event to preview practical Claude Code workflows. Livestream sessions often telegraph the next wave of builder-facing changes.

OpenAI ships Codex pets with /pet, /hatch, and custom companion creation

Codex is getting an ambient UI layer that stays visible while you work. The custom hatch flow suggests OpenAI wants the app to feel more persistent and personal.

Codex /goal keeps a long-running task alive for 11 hours and scores 66 on a senior engineer bench

The long-running task stayed active for hours and still produced a credible benchmark score. That is a strong sign /goal is meant for real project-scale work.

GPT-5.5 stays coherent for 9 hours when given AGENTS.md, MEMORY.md, and PLAN.md

The model only held together because the repo had clear instructions, memory, and a plan file. That is a strong reminder that harness design now matters as much as raw model quality.

Pi v0.71.1 raises throughput 66% by sending only context deltas over cached WebSockets

Only sending new context cuts a lot of transport waste out of long sessions. For agent UIs, network efficiency is now a real performance lever.

Shopify builds a Hermes Agent skill for products, orders, inventory, and fulfillment

Shopify is exposing real commerce operations to an agent through a purpose-built skill. That is the kind of workflow integration enterprise teams actually need.

Gemini lets users put their most-used chat topics into notebooks

Notebook organization gives Gemini a more structured memory surface for recurring topics. That should make it easier to keep long-running work separated and findable.

AskVenice shows how to build AI micro-dramas with character consistency

The guide focuses on the hard part of AI video, keeping characters consistent across shots. That makes the workflow much more usable for serial content.

agent-install installs skills and MCPs across 52 coding agents through API and CLI

One installer for skills and MCPs across many agents is a big DX simplifier. It reduces the fragmentation tax teams pay when they adopt multiple coding agents.