Explore what's new in AI

Paper Shaders goes open source and drops the restrictive wrapper license

The shader library is now free to reuse, resell, and wrap in plugins or Figma tools. That opens a clean asset layer for apps built on visual effects.

LangChain shows recursive language model workflows inside Deep Agents

The tutorial walks through recursive agent loops in Deep Agents and shows how to structure them in practice. That is useful if you are building longer-horizon agents instead of one-shot prompts.

Pauliusztin's context layer diagram separates memory from harnesses

The diagram splits harnesses from unified memory, MCP servers, knowledge graphs, and embeddings. The point is that the context layer is the real moat, not the harness.

Anthropic says Claude now writes more than 80% of merged code

The chart says the share of merged code authored by Claude has climbed sharply. That is a very direct measure of how far internal agent adoption has come.

MCP Server Architecture Patterns paper maps five server shapes and four anti-patterns

The paper separates resource gateways, orchestrators, session servers, aggregators, and adapters. It is useful because tool design matters more when an LLM is choosing the tools.

QuiverAI's Arrow 1.1 generates structured SVG assets through Vercel AI Gateway

Developers can now generate editable SVG assets through the SDK and centralize billing. That is a nice fit for design-heavy workflows that need reusable graphics.

LiteResearcher pitches scalable agentic RL for deep-research tasks

The framework is aimed at training research agents rather than generic chatbots. That makes it relevant for anyone trying to scale deep-research behavior with RL.

SkillBench profiles agent sessions to find token-efficiency bottlenecks

The pitch is to turn traces into a skills profile, then use that to pinpoint where agents waste tokens. That could make delegation and coaching much more measurable.

OpenWiki keeps repo docs up to date for agent consumption

It builds docs from a codebase, updates them as the repo changes, and supports Q and A over both docs and source. That makes it a handy memory layer for teams with large repos.

Bloome runs Claude Code, Codex, and DeepSeek in one group chat

Each agent builds its own version, argues about the logic, and the final site stays in one thread. It is a practical multi-agent workflow, not just a screenshot.

Hermes Agent clears every P0 and P1 issue after an all-nighter

The team says the repo now has zero open P0s and P1s after a week and a half of cleanup. That is a rare public example of disciplined maintenance on an agent project.

v0 adds Nano Banana 2 Lite for image generation in the prompt bar

Turning image generation on inside the prompt bar makes the model part of the normal design flow. That is a nice sign that multimodal output is becoming default, not separate.

Natalia Zarina says Codex built her a custom CRM overnight

She uses Codex for long unsupervised runs, custom tools, and family-care software. It is a strong example of agentic work replacing app stitching.

lgrammel shows AI SDK tools running inside local Docker sandboxes on Mac

The pattern runs shell tools in an isolated container while keeping the host clean. That is a useful safety and reproducibility setup for local agent workflows.

fal's TRELLIS.2 turns 3D asset folders into LoRA adapters

You can fine-tune shape and texture from a folder of assets, then generate matching GLBs. That is a practical way to make 3D generation more style-consistent.

ctatedev says agentic coding and review is going chatless

The argument is that logs, traces, and context should become the UI, with dashboards showing goals, blockers, and outcomes. That shifts agent ops away from chat threads.

Gemini Omni Flash and Seedance workflows edit videos with prompt-level control

The demos show object swaps, environment changes, and morphing transitions while keeping character consistency. That makes video editing feel much closer to prompt iteration than timeline surgery.

Vercel adds dry-run deploys so agents can catch issues before shipping

The dry-run step lets agents validate work before a real deployment lands. That is a simple but high-value way to cut blast radius and review cost.

Codex uses computer use to improve OpenClaw iOS from Twitter feedback

The workflow pulls in social feedback, makes a first-pass change, and even captures before and after screenshots without a GitHub API. That's a neat computer-use workaround.

OpenUI Cloud adds production guardrails for generative UI

OpenUI Cloud is trying to cover the parts that usually get rebuilt later, validation, fallback behavior, versioning, and audit trails. That is the difference between a demo and a production system.

Vercel open-sources konsistent to enforce TypeScript conventions

The same linter used by AI SDK and Chat SDK is now public. It gives teams a shared way to keep agent-written APIs aligned with repo conventions.

Anthropic previewed Claude Science as a local research desktop app

The preview points to a desktop research workbench with local hosting, specialists, and scientific databases. That is a stronger shape for serious literature workflows than a plain chat window.

Google and Anthropic move retrieval into standalone agent services

Both labs are pushing retrieval out of the app and into a service agents can invoke. That helps with stale indexes, shared connectors, and iterative search loops.

Matt Pocock provisions podcast infra in 3 minutes with Gemini and Vercel

The workflow spins up an API key, deployment, and blob storage almost immediately, then falls back cleanly on known errors. It is a good example of agentic bootstrap with guardrails.

Vercel and Shopify rebuild Hydrogen as an agent-first runtime-agnostic framework

The new preview aims at agent-first commerce apps that run anywhere JavaScript does. Runtime-agnostic architecture should make deployments and experimentation less brittle.

Hermes Agent v0.18.0 ships a Judgement Release and new updater

The release adds a new milestone and a one-command updater. That makes the agent stack easier to roll forward as the repo keeps moving.

OpenCode 2.0 hotreloads skills without busting cache

If an agent generates a new skill, the runtime picks it up immediately without busting cache. That makes long-lived coding workflows much easier to iterate on.