Each half of the project landed on the right side of that trade. Proper filing and detailed metadata work can't survive skipped steps. Analysis is judgment work anyway, and I always check what it did.
My longest-running AI agent has never once told me a job was done when it wasn't. It was built the old way: blocks and nodes. It's the intake half of my document archive: a Google Apps Script renames and files every attachment from my email. Meanwhile, a MindStudio workflow
@stevenbjohnson Thanks for sharing this story, especially the bit about the weird little robot. And thanks for NotebookLM! It is a wonderfully useful and easy to use product.
Researchers tricked GitHub's new AI agent into leaking private data with one word.
Noma Labs alerted GitHub that its recently launched Agentic Workflows leaked data out of its private repository.
Their recommendations for builders/AI security officers are worth reading in full. It read as familiar territory from security discussions in our OpenClaw working group.