Explore what's new in AI

Ghostty keybinds jump between agent messages with OSC 133 markers

A simple OSC 133 keybinding lets Ghostty jump between prompt segments and agent messages. That makes long Codex or pi sessions much easier to resume.

Qdrant and Neo4j demo a biomedical GraphRAG copilot

The demo combines vector search with graph reasoning for PubMed workflows, so answers can carry provenance instead of just text.

SonarQube Agentic Analysis plugs CI context into Claude Code

The MCP beta lets Claude Code and Cursor reuse full CI context mid-edit, so teams can fix issues before PRs land.

Hugging Face releases a 10k-query trace dataset for coding agents

The dataset gives tool makers a way to study how coding agents actually use tools, which is useful for evals and harness design.

FastMCP 3.2 adds MCP Apps support for interactive tools

FastMCP 3.2 pushes MCP beyond basic tools into interactive app surfaces. That matters if you want agents to ship richer UI-like experiences over the same protocol.

Logos previews a first-principles physics discovery system

Logos is framing itself as a discovery engine for physics, with early results meant to be automatically checkable. That is a different bet from generic agent demos.

Venice adds 1080p Seedance 2.0 video generation

Venice says Seedance 2.0 can now generate 1080p video, which makes the workflow more useful for real clips instead of only low-res demos.

OpenAI's Codex leak shows draggable desktop avatars

Codex avatars point toward a more personality-driven desktop agent, but the real question is whether users want that much presence.

gog 0.13 adds Gmail forwarding and safer no-send controls

gog keeps moving beyond email automation into a fuller workspace agent, with safer send controls and richer Office-style outputs.

fal ships Tripo3D H3.1 and P1.0 for image-to-3D generation

fal's Tripo3D launch separates hero-quality geometry from production-ready smart meshes, giving teams faster 3D asset output.

ComfyUI's HY 3D 3.0 nodes turn images into editable 3D assets

The node turns single or multi-view images into high-fidelity meshes fast enough for product mockups and game asset iteration.

VS Code lets double-click select bracketed strings and blocks

Double-clicking brackets or strings is a tiny VS Code trick, but it saves a surprising amount of time during quick edits.

Matt Pocock splits LLM work into grilling first, drafting second

Pocock's pattern keeps LLMs from drafting too early, which avoids bloated context and forces a better spec before generation starts.

Google Chrome brings Gemini to more Asia-Pacific countries

Chrome's Gemini rollout is going regional, bundling summarization, image edits, Maps, and Gmail into the browser tab in APAC.

Google DeepMind forms a coding-model strike team under Sergey Brin

The report suggests Google is treating coding quality as a strategic gap, with Brin and DeepMind leadership directly in the loop.

Google's CoDaS agent runs biomarker discovery end to end

CoDaS stitches data prep, hypothesis generation, validation, and reporting into one multi-agent pipeline for biomarker discovery.

Skill-RAG routes failed retrievals to the right repair strategy

The paper turns retrieval into a typed failure-repair loop, so RAG can ask when to retrieve and how to recover instead of retrying blindly.

Exa Deep Max pairs parallel search with a 20x faster latency profile

Exa says parallel tool calls and denser returned text cut latency by 20x while still topping agentic search evals.

MCPorter 0.9.0 turns MCP servers into a TypeScript CLI

The new release smooths MCP wiring from TypeScript, while tougher shutdowns and OAuth fixes make CLI integrations less flaky.

VS Code adds a dedicated shortcut for block comments

One shortcut to comment or uncomment blocks removes a small but constant editing tax during refactors and quick code cleanup.

Cursor CLI adds /debug and /btw for terminal agents

Cursor's CLI now exposes fast bug triage and side questions in-terminal, making long-running agent sessions easier to inspect and steer.

Google's Magika scans files by content, not by extension

Magika catches disguised malware by reading file contents, so teams can scan pipelines locally without trusting extensions or heavyweight servers.