Explore what's new in AI

Gemini now creates Google Docs and Sheets on mobile

The app can draft real docs and spreadsheets from chat, pushing Gemini closer to a workspace editor that makes artifacts.

Sakana's Conductor learns to orchestrate frontier model teams

A 7B controller learned when to route, plan, and verify across models, and it beat single workers and older multi-agent baselines.

MCP Agent Mail's Rust rewrite stabilizes multi-agent mail handling

The new stack holds up better under 15 plus agents, which is a strong signal that the rewrite finally paid off in reliability.

Microsoft's VS Code Learn course teaches custom agents, skills, prompts, and hooks

It is a practical path for shaping AI behavior inside the editor, which is more useful than treating agents as a black box.

c_valenzuelab's "The Cost of a Frame" argues AI video is collapsing production value

The essay says cinematic-looking output is getting cheap, so taste, direction, and what to make become the scarce parts.

Cline Kanban v0.1.64 adds remote servers and per-task model choice

You can run the board on a remote machine and assign different models to different cards, which fits parallel agent work well.

OneManCompany treats agents as recruitable talents in a labor market

Instead of fixed teams, it recruits portable agent identities per task, and that dynamic setup beats prior SOTA on PRDBench.

Ramp says its internal coding agent now writes 60% of merged PRs

That is a meaningful adoption metric, and it suggests structured product context can move agents from helper mode into throughput.

Parallel adds x402 payments so agents can pay for search at runtime

Runtime micropayments let agents buy search and extraction on demand, a practical path toward usage-based web access.

Perplexity ships Sonar 2 as its in-house web search model

The model is tuned for fast grounded answers, which matters if you care more about search latency than flashy chat.

Agentic World Modeling maps predictors, simulators, and evolvers

The paper gives builders a shared vocabulary for world models across physical, digital, social, and scientific settings.

BridgeBench puts GPT-5.5 below GPT-5.4 on real-world vibe coding tasks

Despite the hype, GPT-5.5 lands behind GPT-5.4 here, showing that more general intelligence does not always mean better dev UX.

OpenCode adds automagical Zed support for terminal-driven coding

The update makes it easier to bounce between terminal workflows and the Zed editor without manual glue.

GPT-5.5 plus GPT Image 2 assembles watch components in one Codex pass

The demo shows multimodal agent workflows can generate and align complex visual parts in one shot, useful for design-heavy tasks.

GPT-5.5 jumps from 36.6% to 73.7% on Matharena

The leap on fresh olympiad-style problems suggests a real reasoning gain, not just benchmark overfitting or prompt luck.

pi.dev launches a new website and lets you update Pi from the CLI

The update pairs a cleaner site with a direct `pi update` path, which makes the terminal harness easier to adopt and keep fresh.

Qwen 3.6 Max Preview lands on OpenRouter with 262K context and agentic coding focus

The release adds a huge context window and tool-use emphasis, making it a serious option for long-horizon coding and agent tasks.

Arena ranks GPT-5.5 across code, docs, search, vision, and math

GPT-5.5 lands strong but uneven across modalities, with a real code jump and solid search performance, but not a clean sweep.

Simon Willison runs Microsoft VibeVoice locally on an M5 MacBook

A 1 hour recording transcribed in about 9 minutes on-device, showing local speech models are getting practical for serious notes.

GPT-5.5 will reach a 1M-token Codex context window, up from 400K

If the issue lands, longer agent sessions and giant repo edits get much easier, which matters for real refactors and long tasks.

Hermes Agent turns native vision on by default for supported models

That cuts setup friction, because you only need an auxiliary vision model when the main model truly lacks multimodal support.

Cognition says Mercedes-Benz is deploying coding agents across global teams

This reads like a real enterprise roll-out, not a pilot, and it suggests agents are moving into large industrial orgs.

Gemma 4 and Pi run a fully local coding agent with read, write, edit, and bash

The setup shows a practical offline agent loop, with a local model server and a tiny tool set that can write and run code end to end.

Matt Pocock builds a /diagnose skill that reproduces bugs and tests fixes

The workflow wraps a bug in an AFK loop, shrinks the repro, then tests hypotheses, which is a strong pattern for agentic debugging.