Explore what's new in AI

Sakana AI launches Fugu beta, a multi-agent orchestration system

The beta turns multi-agent coordination into a product, which is useful because orchestration is where many stacks fall over.

Azure blocks prompts that contain the word System

A single reserved word can break your prompt plumbing, which is exactly why provider-specific prompt tests matter.

NanoClaw powers a diplomat's 24/7 second brain on a Raspberry Pi 5

This is a concrete proof point for local, open-source second-brain setups in high-stakes knowledge work. A Pi 5 is enough if memory is good.

GitHub Next shows Ace, a multiplayer agent environment with shared context

Ace tries to fix the solo-agent problem by giving humans and agents the same sandbox and shared context, so they can plan together.

Sakana's TRINITY routes tasks across thinker, worker, and verifier roles

A tiny coordinator plus evolutionary search beats single-model baselines on coding and math, which is a useful orchestration pattern.

Google Scholar Button and PDF Reader cut paper triage friction

The combo gets you from citation to PDF with fewer clicks, which is exactly what you want when skimming lots of papers.

Rat turns VS Code into a shared REPL workspace with code cells

Persistent REPL cells and shared kernels make agent collaboration in VS Code feel much less like context roulette.

Claude Code now prints session IDs and syncs /login across active sessions

Small fix, big UX win. Account switching is less brittle when one login updates the rest of the running sessions automatically.

Matt Pocock says explaining why you compact improves the result

If you tell the model what to keep and what to drop before /compact, it preserves the right context and throws away less useful clutter.

Birdclaw stores Twitter archives locally and backs them up to GitHub

It solves the annoying part of owning your social data: archive restores, daily bookmark imports, and automatic backups without relying on the API.

Browser Use Cloud lets you spin up autonomous browser agents

Moving browser agents to the cloud makes long-running automation more practical, especially for tasks that need persistent sessions.

wterm adds a Vue component for a 12 KB terminal emulator in the browser

The DOM-rendered terminal keeps selection, search, and accessibility intact while staying tiny. Good if you want terminal UX inside apps.

Codex users hand OpenAI a long bug list on desktop UX and agent flow

This is a dense backlog of real friction, from update prompts to dead buttons and missing worktree state. Great raw product feedback.

Yacine MTB wires Codex completions to a desktop notification sound

Tiny but useful workflow polish. A sound on completion keeps fast agent loops visible without babysitting the terminal.

OpenAI's image model changes enough that rerunning prompts can pay off

The same detailed prompt can land very different outputs, so important image work should be sampled a few times instead of trusted from one shot.

Sam Altman says the internet should work for people and agents

The tweet points to a bigger UI and protocol shift, where OSes and web apps need to serve humans and autonomous agents equally well.

OpenClaw's clawsweeper closes 27k issues and 30k PRs since December

This is a real workload metric for agentic repo maintenance, not a demo. The bot is now clearing backlog at industrial scale.

Zed lines up bookmarks, code lens, and markdown rendering upgrades

The next few weeks add practical editor ergonomics, especially for agent-heavy workflows that live in markdown and commit history.

YC lays out how to build a company around AI from day one

The doc pushes queryable orgs, fewer human middleware layers, and token-max operations. It's a snapshot of the startup operating model shift.

OpenCode scores highest on a social-science reproduction benchmark

The result suggests harness choices and extra search work can matter as much as the base model when reproducing research tasks.

Blacksmith lets Codex burst test suites on 32 vCPU instances

The practical win is simple: offloading local tests to a CPU box can speed up agent loops without burning your laptop. That matters for long validation runs.

Small Harness puts its first official repo release in public

The repo goes open source and gives harness builders a concrete starting point instead of an opaque hosted setup. Useful if you want to inspect and fork the stack.

Percy Liang's talk on code synthesis for better agentic decision making

The talk connects AutoHarness and code world models to a broader idea: better code synthesis can improve how agents decide, not just what they write.