Explore what's new in AI

Zed 1.0 ships a complete Rust editor across Mac, Windows, and Linux

Zed treats 1.0 as the point where most developers can actually live in it, backed by a Rust rewrite, cross-platform support, and ongoing weekly shipping.

OpenAI schedules DevDay 2026 in San Francisco for September 29

OpenAI has locked in the date and city for DevDay 2026. That gives developers a clear checkpoint for upcoming API and platform announcements.

Gemini generates Docs, Sheets, PDFs, and Office files directly in chat

Gemini can now produce PDFs, Docs, Sheets, and Office files straight from chat, cutting out templates and copy-paste for routine docs.

LangChain's DeepAgents Deploy CLI launches agents from markdown and config

The open-source CLI wires up and deploys agents with minimal code, aiming to make production rollout mostly declarative.

Phil Schmid publishes a Gemini Deep Research developer guide

The guide shows async research, approval loops, chart generation, and streaming updates for production Deep Research workflows.

Cursor's /multitask splits long agent work across async subagents

Cursor can offload inspect, plan, code, and test steps into side agents, keeping the main chat cleaner and more focused.

Google DeepMind and Korea team up on AI for scientific discovery

The partnership centers an AI campus and joint research to accelerate life sciences, climate, and broader national AI capability.

Latent Agents distills multi-agent debate into one model

The paper keeps debate gains while using up to 93% fewer tokens, and even exposes agent-like circuits in activation space.

Hermes Agent automates Stripe refunds and verifies the result

A live demo shows Hermes finding a customer in Stripe, issuing a refund, then verifying the payment is no longer refundable.

Sakana and SMBC deploy a multi-agent proposal generator into banking workflows

Sakana and SMBC say multi-agent proposal drafting can cut weeks of banking prep down to hours. The rollout shows real enterprise value, not just a demo.

Virtual Puppet Theater uses voice and hand tracking to stage AI puppets

The demo combines hand tracking, voice, and AI-controlled props into a live browser stage. It is playful, but it also shows how multimodal interaction can feel natural.

ARIA turns machine manuals into technician work orders with prior fixes

ARIA reads machine manuals and turns past fixes into technician work orders. That is a concrete managed-agent pattern for maintenance, not just generic support chat.

MaestrIA photographs home damage and drafts a repair plan plus local quote

MaestrIA photographs home damage, estimates local repair costs, and drafts a message to a nearby tradesperson. It turns multimodal diagnosis into a practical workflow.

Maieutic makes students explain the plan before they can write code

Maieutic disables coding until students explain the plan in plain language. That turns an AI editor into a reasoning tutor instead of a syntax crutch.

Wrench Board reads schematics and draws step-by-step repair guidance on the board

Wrench Board turns schematics into step-by-step repair overlays on the board itself. That's a strong pattern for multimodal technical troubleshooting.

Google Research's empirical research assistance automates code search inside science loops

Google says researchers are already using Empirical Research Assistance in epidemiology, cosmology, and neuroscience. The workflow turns scorable problems into sandboxed code loops.

spawn-agent exposes Claude Code, Codex, Cursor, and Pi through one API

spawn-agent gives one API for Claude Code, Codex, Cursor, Pi, and other local agents. That could reduce harness glue when you want to swap runtimes.

Hermes Curator prunes stale agent-created skills on a weekly cadence

Hermes Curator adds a maintenance loop for agent-written skills, pruning, consolidating, and archiving stale ones before the skill pile gets unmanageable.

Hermes Agent adds a pretext skill for DOM-free text layout and geometry

Hermes Agent's pretext skill focuses on DOM-free text layout, text reflow, and kinetic typography. It is aimed at creative browser pieces and design prototyping.

OpenCode 2.0 makes distributed sessions embeddable with automatic resync

OpenCode is turning workspaces into embeddable distributed sessions, with state that resyncs after downtime. That makes remote agent workflows much harder to break.

Portless adds Tailscale sharing for local dev servers and monorepos

Portless can now expose local servers to your tailnet or the public internet, which removes a lot of friction from remote demos and multi-repo dev work.

Matt Pocock open-sources Sandcastle as a software factory for agents

Sandcastle packages an agentic software-factory workflow into a repo you can reuse. That matters because harness design is becoming a real engineering surface.

Agentic Harness Engineering lifts Terminal-Bench 2 by making harness edits observable

The paper treats harness edits like observable, revertible code and lifts Terminal-Bench 2 pass@1 from 69.7% to 77.0% in ten iterations.

OpenAI shows Codex drafting readouts from Slack, Gmail, and Calendar changes

The demo pushes Codex beyond code into support work, like comparing inboxes, calendars, and exports before writing a readout.

OpenAI publishes GPT-5.5 prompt guide for shorter, outcome-first prompts

OpenAI says GPT-5.5 works best when you describe the outcome, constraints, and evidence, then let the model choose a short solution path.