Explore what's new in AI

Browser Use tops Odysseys on long-horizon web workflows

Multi-hour browser work is the real stress test for agents. Strong results here suggest the stack is moving past short single-step demos.

TokenPilot cuts agent costs with cache-friendly context management

The paper argues that cheaper agents need stable memory, not just shorter prompts. Preserving cache layout while evicting history delivers big savings.

Neo MCP hands long AI and ML loops off from Claude Code

The MCP bridge lets a coding agent delegate long-running experiments, evals, and debugging to a worker that returns code, metrics, and execution trails.

Codex runs a 4-day autonomous data-management project and ships apps

This is the clearest sign yet that coding agents are moving from chat to multi-day work. The human is now supervising outcomes, not every turn.

Corey Ganim turns live AI workflow help into a recurring service

The model works because it sells implementation, not advice. One or two live calls a month can turn into steady MRR if each session ships something useful.

Doodlestein traces Rust build stalls to btrfs writeback bursts

The issue was not CPU, it was writeback. Tuning dirty_bytes and linker priority kept the machine responsive while builds were running.

The Codex Chrome extension automates routine browser tasks

The practical win is letting Codex handle repetitive browser work so humans can stay on higher-value tasks. It is the kind of feature that quickly becomes habitual.

agent-browser MCP wraps browser automation for agent loops

The MCP wrapper gives agent builders a cleaner way to plug browser control into their loop. That matters if your agent already works in a GUI-heavy workflow.

Gemma 4 gains 5x throughput in a crowd-sourced agent collaboration

Over 100 agents traded messages and pushed throughput from 100 tok/s to 500 tok/s. It is a nice example of coordination and model tuning together.

FarzaTV ships an AI tutor that draws on your screen

Drawing directly on the user interface turns the model into a visual instructor, which is a practical pattern for step-by-step teaching and debugging.

Firecrawl lets you search, scrape, and parse PDFs free without an API key

It lowers the barrier to trying Firecrawl on real web tasks, and the MCP plus CLI support make it immediately useful inside agent workflows.

Google ships Android 17 with bubbles, screen reactions, and safer flows

Android 17 folds in a few useful AI-adjacent behaviors, plus stronger safety features. The screen reaction and compact window changes are the most developer-visible ones.

Anthropic keeps Claude subscription programmatic use on plan

The reversal matters for builders who script Claude. It keeps subscriptions usable for lightweight automation instead of forcing every workflow onto API billing.

Claude Code 2.1.179 fixes drops, sandbox bugs, and stuck jobs

The changelog patches mid-stream disconnects, broken sandbox globs, and phantom running jobs, exactly the sort of reliability issues that hurt long sessions.

PyLate regularizes ColBERT models to restore efficient ANN search

Modern multi-vector retrieval broke older ANN tricks, and this fix shows how training regularization can bring them back for practical search stacks.

GitHub Copilot app becomes a home base for triage, telemetry, and prep

GitHub is showing Copilot used beyond PRs, which is a useful signal that agent tooling is spreading into everyday operations.

Claude Code paper says the harness matters more than the agent loop

The paper makes the architecture point clear, a tiny loop can work if the harness handles tools, memory, permissions, and recovery. That's where production complexity lives.

ComfyUI rebuilds sky replacement into a full animated VFX workflow

This is a practical creative pipeline, not just a prompt. It takes footage in, rewrites the sky, adds motion, and outputs a finished shot with metadata intact.

Electric Agents 0.6 adds spawn, fork, wake, and schedule primitives

The release pushes the idea that the agent is the log. Durable streams and local or remote runners make long-lived agent systems much easier to build.

Rivet's Effect SDK gives actors durable state, websockets, and tracing

The SDK treats actors as long-lived processes with persisted state and observability. That makes it easier to reason about services that sleep and resume.

OpenAI says heavy ChatGPT users save more than 10 hours a week

The report shows AI impact compounds with depth of use, and the biggest time savers are also the heaviest credit users. That's useful for rollout and budgeting.