Explore what's new in AI

Anthropic's Dario Amodei says software could become cheap enough to upend app economics

He says Anthropic engineers already let Opus write most of the code and then edit it. That is a useful signal for how frontier labs see software work shifting.

TanStack Router adds more control over dynamic route params

params.parse now lets you accept or skip a candidate and resolve overlapping dynamic params with explicit priority, which is handy for complex route trees.

Adala automates data labeling with autonomous agents

Adala turns labeling into a learn-by-doing loop over ground-truth examples, then swaps models underneath the same skill. That makes cheaper labeling and student-teacher setups easier.

Heyclicky launches a voice-first agent UX with proactive nudges

Heyclicky's new agent UX tries to live in the user's workflow instead of the terminal, using voice and proactive nudges to interrupt only when useful.

Cursor's weekly insights dashboard shows 214 active chats and 16k active minutes

Cursor's weekly insights dashboard shows how much hidden usage the app accumulates, and why teams want built-in telemetry over ad hoc scripts.

Squid turns Claude Code into a five-agent spec-to-PR pipeline

Squid turns Claude Code into a five-agent assembly line, spec to plan to tests to PR review, with only two human gates in the loop.

Parlant adds criticality levels so agents can weight instructions

Parlant's criticality levels let teams mark some instructions as non-negotiable and others as soft guidance, a better fit for production agents.

WeRead's beta skills package exposes bookshelf, stats, exports, search, and recommendations

WeRead's beta skills make the reader app much more agent-friendly, with actions for bookshelf, stats, note export, book search, and recommendations.

Finance agent workshops say OCR and HITL review are required for document context

This workshop says finance agents need rigorous OCR, eval checks, and human-in-the-loop review, because a single number error can cascade downstream.

Model-Adaptive Tool Necessity finds the knowing-doing gap in tool calls

This paper isolates a classic agent failure mode, models often know a tool is needed but still fail at the action step, which explains many prompt ceilings.

Codex can control the terminal it is running in from the CLI

A CLI trick lets Codex operate the terminal it lives in, which is useful if you want a single shell session to host both the agent and its commands.

Claude Code's YAML frontmatter parser accepts invalid YAML in agent files

Claude Code's agent-file parser accepts invalid YAML, so tooling around it may need to mimic the same broken edge cases if you want compatibility.

δ-mem adds online memory for long-term assistants without extending context windows

δ-mem aims to give assistants durable recall without ballooning context windows, which matters for long-lived agents that need history but not full transcripts.

pi-ai 0.74.1 adds image generation, Together AI, and Windows ARM64 binaries

pi-ai 0.74.1 adds image generation, a new provider, and Windows ARM64 artifacts, plus a long bug-fix pass that makes the SDK more shippable.

Gemini 3.5 Flash builds a full Ubuntu GNOME web OS in one shot

A one-shot Gemini demo built a full desktop-like web OS with file manager, terminal, and window controls, showing how far code generation has moved.

Google Cloud Console leaks Claude Mythos and Gemini 3.2 Flash-lite-live

Cloud console quota screens exposed new Anthropic and Google model names before launch, a reminder that provider UIs often leak the roadmap first.