Best of the Week: Anthropic's Agent Hours Report and 4 More AI Tools

A lot shipped this week, and most of it points the same way: agents are doing more of the work, and the people running them are the bottleneck. These are the five Shorts I'd send a friend who builds with AI, in the order I'd send them. Each one links to the full breakdown with sources.

Anthropic's AI Agents Log 6x Their Humans' Hours

Anthropic studied its own workforce and found that by August, the agents its employees direct were logging six times the hours of the people running them, while interactive sessions sat around 80% idle waiting for human input. The report calls the new work "meta-work": writing specs for multi-agent runs, having one agent review another, and building monitors for agent pipelines. If you run coding agents, this is the clearest picture yet of where your time goes next.

Full breakdown with sources: 500xengineer.com/blog/anthropic-agent-hours

Claude Found a Year of Security Bugs in One Week

Anthropic and Theorem had Claude write machine-checked proofs that 239 entry points across 123 open-source libraries don't crash on any input up to 64 MB, and the work turned up hundreds of bugs along the way. Anthropic is clear that "formally verified" does not mean "bug-free," since a proof is only as good as its specification. Open-source maintainers and anyone who cares about security should read the limits section before the headline.

Full breakdown with sources: 500xengineer.com/blog/claude-formal-proofs

NVIDIA Built a Jail for Your AI Coding Agent

OpenShell is NVIDIA's open-source runtime that puts each agent in its own sandbox, checks every network connection against a policy, and never hands the agent your real credentials. Risky policy changes wait for human review. If your coding agent runs in your terminal as you, it can read your .env, and this is a concrete way to stop that.

Full breakdown with sources: 500xengineer.com/blog/nvidia-openshell

One Developer Rebuilt 7 Adobe Apps With Claude

Brandon Thomas released seven free, open-source Rust apps that reimplement Adobe products, from Photoshop to InDesign, with Claude writing much of the code, though he says they're "not anywhere close to ready." Alongside it, the REA toolkit lets coding agents take apart an app and explain how it works. Builders who maintain old software or study competitors should see how cheap copying behavior has become, and where the legal questions are still open.

Full breakdown with sources: 500xengineer.com/blog/artcraft-rea-app-cloning

Why Your AI Forgets: The Context Window, Explained

Your AI only sees what fits in its context window: the oldest turns of a long chat fall out first, and facts buried in the middle of a long prompt often get missed. The fix is a few habits: put key instructions at the start or the end, restate key facts in long chats, and start a fresh chat for a new task. Anyone who writes long prompts gets better answers from this one.

Full breakdown with sources: 500xengineer.com/blog/context-window