Explore what's new in AI

Anthropic says Claude now writes more than 80 percent of production code

The disclosure suggests Claude has moved from suggestion engine to delegated engineering work. The remaining human value is picking problems and judging results, not typing every line.

OpenAI's Build iOS Apps plugin lets Codex test apps inside the loop

The plugin closes the loop by letting Codex view and test iOS apps instead of guessing from code alone. That should make mobile agent workflows much less brittle.

NotebookLM adds source attribution and exports to PDF, DOCX, Markdown, and text

Source attribution and richer exports make NotebookLM artifacts easier to audit and reuse. That's a practical upgrade for teams turning notes into docs and reports.

Google tests Gemini Troubleshooting Mode with step-by-step widgets

The mode walks users through problems with guided explanations and interactive widgets. It's a useful sign Gemini is moving from chat to hands-on problem solving.

Google Gemini Live turns image edits and synced text into a real-time loop

Gemini Live now creates and edits images while keeping typography and motion tied to the scene. That makes it feel more like live creative editing than batch generation.

Hugging Face shows its CLI uses 6x fewer tokens than raw API calls

On about 1,000 graded runs, hand-rolling curl or SDK calls used up to 6x more tokens and failed more often. The result argues that good CLIs are cached intelligence for agents.

Google shows proactive multimodal agents for intelligent eyewear at CVPR

Google's eyewear demo turns first-person context into actions, and the linked Ego2Web benchmark makes the task concrete. It's a strong signal that wearables plus web agents are converging.

DSPy says prompts should be backed by typing, tests, metrics, and optimizers

DSPy treats prompts as imperfect instructions and pushes builders to validate them with types and metrics. That's a practical antidote to prompt folklore.

OpenCode's /move names worktrees from your prompt for frictionless parallel work

OpenCode now auto-generates useful worktree names from the task prompt, which removes a surprisingly annoying bit of branch setup. Small defaults like this make parallelism feel natural.

Amp's deep and rush modes finish long tasks up to 40% faster

Amp says its rebuild cut first-token latency by 87% and end-to-end time by up to 40% on long tasks. The win comes from fewer hops and more durable execution around long-lived websockets.

browser-wall controls multiple cloud browsers from one CDP endpoint

browser-wall lets one agent manage several cloud browsers at once, with profiles and proxies preloaded. That's a handy pattern for parallel browsing, QA, and anti-bot workflows.

OpenSquilla's MetaSkill lets agents discover and compose their own skills

MetaSkill tries to replace hand-written workflows with a self-organizing skill layer. The clever bit is routing easy tasks to cheap models and reserving heavier models for the hard parts.

Matt Pocock defines primary and secondary sources for context engineering

Primary sources are raw truth, secondary sources are summaries and docs. The framing is useful because compaction makes lossy summaries inevitable, so source hierarchy matters.

ChatGPT memory and Codex can be bridged through Obsidian notes

A practical workaround is emerging: write ChatGPT memory summaries to Obsidian, then have Codex read them back. That gives you one shared memory layer across tools today.

OpenAI Codex users are turning the tool into inbox triage, automations, and note taking

People are using Codex for more than code: inbox zero, scheduled automations, and notes to Obsidian. That's a strong sign it is becoming a general ops assistant.

OpenAI redesigns API navigation to make the platform easier to scan

OpenAI's API console got a new layout that makes finding docs and settings easier. This is the kind of small DX polish that cuts support friction and setup time.

Sakana AI launches RSI Lab to build recursive self-improvement systems

Sakana is formalizing two years of self-improvement work into a lab focused on open-ended, sample-efficient AI research loops. That's a clear bet on agents improving their own builders.

Google's agentic RAG uses a Sufficient Context Agent to keep searching

Google's new RAG flow rewrites queries, fans out searches, and checks whether it has enough context before answering. That shifts retrieval from one-shot lookup to iterative reasoning.

ChatGPT lets you send email directly from writing blocks

ChatGPT can now draft, tweak, and send email inside the conversation, cutting a round trip through the inbox. It turns the web app into a lightweight workflow surface.

Anthropic doubles Claude Cowork usage limits for bigger tasks

Anthropic is temporarily doubling Cowork limits, which makes long, messy tasks easier to delegate for a month. If your work spans hours, the quota bump matters.

OpenAI Codex adds settings search, persistent state, and themed profile cards

Codex is smoothing the rough edges: searchable settings, state that survives restarts, and profile cards that match your theme. Small UX fixes can make a tool feel much more usable.

Tencent WorkBuddy turns a PC productivity agent into MCP-connected expert teams

WorkBuddy routes tasks to specialized sub-agents and external apps, so reports, docs, and desk work can all happen from one chat. That's a concrete enterprise agent workflow.