When you swap your model, what will you keep?

The model is turning into a commodity. What makes an agent useful or dangerous is everything we build around it: what it knows, what it can touch, what stops it, and who holds the risk. This week we explore the perimeter from the inside out.

• PRACTICE · Kevin Liao
  Agents Don’t Need Memory. They Need Documentation. (opens in a new tab) · context is a set of documents the agent reads and keeps current, so anyone can open them and check what it believes
• THE STAKES · Michael Catanzaro
  The Era of Software Quality, or the Era of Ostriches? (opens in a new tab) · AI now finds real security bugs faster than volunteers can read the reports: GNOME paid out €183,900 for 71 of 298 bounty reports, then shut the program down because its one maintainer was buried
• DISSENT · Matthew Green
  Is sandboxing sufficient to contain rogue agents? (opens in a new tab) · a useful agent needs tools, data and a way out, so the box always has a door, and a second model standing guard only moves the trust
• THE BILL · Simon Willison
  We’re going to need default hard budget caps on pretty much everything (opens in a new tab) · spending should stop at a cap that is on by default, because the warning email arrives after the money is gone
• WIDER VIEW · Armin Ronacher
  The Tower Keeps Rising (opens in a new tab) · agents removed the friction that kept a team speaking one language about its code, and the building goes on after the shared understanding is gone