Explore what's new in AI

Hyperbrowser extracts a site's design system into the agent workflow

The terminal command pulls a site's design system straight into the coding agent, which is a nice way to stop hand-translating UI style into prompts.

Emdash v1.1.19 fixes SSH, PR diffs, and Windows edge cases

Emdash v1.1.19 tightens the boring parts of remote coding, SSH, diff review, Windows paths, and integrations. Those reliability fixes are what make agent tooling usable every day.

kepano's Obsidian skills teach agents Obsidian file formats

The skills package teaches agents Obsidian's markdown, bases, canvas, and CLI workflows. That is a solid pattern for making agents native to a real app ecosystem instead of generic note-takers.

Perfect HQ says Qdrant multivector search raised match accuracy from 30% to 99.993%

The case study shows multivector retrieval and orchestration can push recruiting acceptance into the mid 90s while staying under latency budget.

Antoine Chaffin says BM25 and ColBERT still dominate deep research traces

The thread argues that late interaction stays strong because query formatting and OOD robustness matter more than many teams expect. Good retrieval still starts with the trace.

dzhng turns failing traces into a prompt engineering loop

Instead of hand-writing prompts, the workflow feeds bad traces into evals and red-green iterations. That is a more reproducible way to improve agent behavior.

OpenAI's Codex goal mode gets SQLite state tracking and lifecycle hooks

Goal state now lives in a dedicated DB and extensions can observe tool, subagent, and turn events. That makes long-running agent work easier to orchestrate.

Linear adds automatic Slack channels for every new project

Projects now get their own channel and updates flow in automatically, reducing the manual ceremony around kickoff and status drift.

swyx builds a skill to harden vibecoded apps into tested repos

After 16 hours and 103 commits, the skill produced a maintainable codebase without changing the app's behavior. That's a concrete hardening workflow.

GeminiApp showcases Gemini Omni creations across multimodal workflows

The showcase suggests Gemini Omni is becoming the catch-all frontend for text, audio, image, and video creation. That broadens the model's role beyond chat.

Google Gemini Daily Brief turns Gmail, Calendar, and Tasks into a morning digest

Daily Brief pulls ahead on the user's behalf and surfaces what matters before the day starts, which is a strong pattern for proactive assistants.

Hark raises $700M to build personal intelligence hardware and models

The company is using a huge round to scale GPUs, hire up to 200 engineers, and push into voice, memory, and computer-use hardware.

Google Pomelli builds brand books and a site from files and images

Pomelli can build a Business DNA profile from files and images, generate brand books, and launch a site in one click. That lowers the bar for fast campaign production.

Google Flow brings agentic video and music workflows to mobile

Google Flow now has iOS and Android apps, plus agentic tools for collaborative video and music workflows. The interesting part is treating creative production as a multimodal agent loop.

Warp hands an agent session off to the cloud and back

Warp can now move a live agent to a cloud runtime and return it later, which makes long-running work easier to carry across devices.

Gemini 3.5 Flash tops APEX-Agents-AA and runs well on RuneScape-Bench

The model ranks #1 on one agent benchmark and also performs well on a game-like benchmark, suggesting Google's latest Flash model is more practical than expected.

ChatGPT can now edit PowerPoint slides directly inside PowerPoint

The beta keeps slides editable instead of exporting a flat deck, which makes AI slide drafting a lot more practical for real work.

Entire CLI bakes Pi into the main agent workflow

Pi is no longer an external plugin, and now connects directly to checkpoints, commits, and session history inside the CLI.

v0 adds one-click PR merge conflict resolution

The agent now merges the base branch and resolves conflicts directly, which trims a common friction point in review-heavy teams.

OpenBMB's MiniCPM-V 4.6 auto-tags fashion product images

A 1.3B model can structure product imagery into business labels, which is useful for preprocessing and catalog ops.

Zed v1.3 adds Terminal Threads and a git history tab

Zed now lets you run Claude or Codex beside built-in threads, which tightens the loop between editing, agent work, and commit history.

Firecrawl ships search, scrape, and interact APIs for agent web access

It packages web search, page cleanup, and login-click automation into one API, a practical baseline for agent web pipelines.