Explore what's new in AI

Hayden Bleasel ships docs inside npm packages to help agents

Shipping docs inside node_modules is a clever way to give agents local context without asking them to scrape the web first.

Figma's MCP server round-trips designs and live pages with agents

The interview shows Figma can turn a web page into canvas and hand designs back to an agent, but review still looks like the bottleneck.

Engram manages agent memory as a background service on Weaviate

Engram treats memory as a managed service, reconciling new facts and preferences in the background instead of just stuffing a vector store.

TripoSplat turns a single image into 3D Gaussians in under 5 seconds

The API turns a single image into high-quality 3D Gaussians fast enough for rapid asset prototyping and layout iteration.

Arrow 1.1 turns illustrations into editable SVG source code

Arrow shifts visual AI from pixels to code, which makes outputs easier to refine, reuse, and plug into real workflows.

Paper's Pen Tool turns sketches into editable SVGs

Paper's Pen Tool lets teams draw, vectorize, and edit SVGs in place, so export fidelity finally matches the canvas.

Zed v1.5 renders Mermaid diagrams faster and more accurately

The release makes Mermaid diagrams render faster and more accurately, which matters if you use docs-heavy workflows inside the editor.

Linear Diffs brings code review into Linear

Reviewers can inspect diffs in Linear, sync decisions back to GitHub, and keep agent-assisted changes in one place.

Lovable says one internal app now handles over half of data queries

A single Lovable app became company infrastructure by combining warehouse connectors, RBAC, dashboards, and natural-language reporting.

OpenCode says 10M developers and 5x TPS justify its caching stack

Baseten's numbers make a strong case that caching and serving work matter: more throughput, lower cost, and 10M developers using the product.

Google DeepMind's science-skills repo packages research workflows

These skills package domain-specific procedures for genomics, biology, and literature search, giving agents reusable scientific workflows.

OpenAI's data agent enriches table context at warehouse scale

The blog shows context assembly, not raw SQL cleverness, is what makes a data agent usable at OpenAI's warehouse scale.

Hyperbrowser's /web turns any website into a web.md map for agents

This is a simple but useful idea, map the page first so the agent knows the structure before it starts clicking around.

Firecrawl Workflows turns repeat web tasks into installable skills

The new workflows let teams turn research, SEO, and cloning tasks into reusable CLI skills, which is a tidy way to standardize web automation.

Hornet builds retrieval for agent workloads, not human keyword search

Hornet is tuned for long, structured agent queries and claims lower serving cost at scale, exactly where legacy search stacks struggle.

Cursor's /loop skill keeps agents running on schedules and terminal output

This is a practical way to turn a coding agent into a recurring watchdog, useful for PR polling, infra checks, and other time-based tasks.

Codex plugins let the model pick and install useful extensions itself

Instead of guessing which plugin you need, you can ask Codex to inspect your repo and workflow and install the right extensions for you.

Avichawla's agent harness splits memory, skills, and protocols

The framing moves intelligence out of the model and into a runtime harness, which is exactly how agent systems are becoming easier to reason about.

Anthropic's analytics blog turns data foundations into an agentic stack

Anthropic’s playbook says reliable analytics agents come from clean warehouse foundations, curated skills, and explicit eval loops, not raw prompting.