Explore what's new in AI

Jason Zhou gives a coding agent a repo map and richer grep hooks

A repo map plus richer grep tracing cut token use by about 50% and give the agent better blast-radius context. Good pattern for large monorepos.

Matt Pocock packages Martin Fowler code smells into a /review skill

The /review skill encodes familiar code smells into an agent checklist. That should make automated review less random and closer to human standards.

OpenClaw says 5% of production traffic runs on Pi and 5% on OpenCode

The update says 5% of production traffic runs on Pi harnesses and another 5% on OpenCode. That's a concrete signal that agent stacks are diversified in production.

Portless lets one local proxy serve multiple TLDs for dev testing

One proxy instance can now serve multiple TLDs locally, so you can test realistic domain routing without extra setup. Small, but handy for dev loops.

Spellbook launches autonomous contract management across the full deal lifecycle

Spellbook is moving from point-in-time review to contract infrastructure across intake, negotiation, renewal, and maintenance. That's a real workflow platform play.

OSWorld2 benchmarks computer-use agents on long-horizon desktop tasks

OSWorld2 pushes computer-use agents onto long-horizon desktop work, which is the right test for real automation. Short demos won't be enough anymore.

DeltaDB records agent edits as ordered deltas instead of commit snapshots

DeltaDB stores every edit with the conversation that produced it, which matches how agents work between Git snapshots. That makes provenance easier to query.

Neural Procedural Memory stores agent know-how as activation steering vectors

The paper argues text memory is not enough for execution skill, so it distills behavior into a procedural layer. Promising for persistent agent know-how.

Eve uses OpenUI to turn chat output into editable dashboard UI

The demo turns a chat prompt into a live dashboard UI, then keeps the agent active while you refine it. That's a useful pattern for generated interfaces.

AI CLI adds text-to-speech and speech-to-text in the terminal

You can now speak prompts and hear replies without leaving the shell. That makes quick agent loops easier when your hands are busy or you're screenless.

MCP Server Architecture Patterns paper maps five recurring server shapes

The taxonomy helps teams stop reinventing the same MCP shapes and decide when to expose resources, orchestrate tools, manage sessions, or adapt workflows.

ASPIRE turns robot practice into a reusable skills library

ASPIRE stores sensorimotor know-how as skills that compound over time instead of retraining from scratch. The team says it can cut transfer compute by 10x.

Matt Pocock's /wizard skill turns third-party setup into an interactive CLI

The wizard turns third-party setup into a guided CLI with saved state, so long provisioning runs are less error-prone. A nice pattern for agent-assisted onboarding.

John Carmack argues code may work better as parsed trees than flat tokens

He suggests binary code could be parsed into context instead of flattened into tokens. That could make long code reasoning more structured than plain text.

fofrAI publishes an Omni API skill for video editing workflows

The skill bootstraps editing, text-to-video, first-frame generation, and prep helpers in one place. That's a ready-made starting point for agentic video pipelines.

Paper adds tokens, pen tool, cursor chat, folders, and Figma paste

Paper's latest update pulls more edit, review, and import work into one canvas. That reduces handoff friction for design teams.

Acti turns the keyboard into an action layer for agent workflows

Acti makes the keyboard an action layer, so you can type a request, hold a key, and fire a workflow without switching apps. Nice fit for agentic shortcuts.

Anthropic frames managed agents as a stable layer over changing harnesses

The talk says engineers are shifting from building models to orchestrating systems around them. That's a good mental model for long-horizon agent work.