Agents need better lanes, proof, and tighter permissions

Made by Mechanical Means shows Claude coordinating dozens of cloud agents to explore and build a homepage redesign without local code or traditional development tools. It’s a concrete orchestration pattern for dividing work across agents while keeping the overall outcome in view — Principle 09.

Cloudflare open-sources Forge, a CI pipeline that previews and validates generated SDKs, CLIs, and docs across hundreds of APIs. Treat generated artifacts as production outputs with automated checks, not as agent work to trust on sight — Principles 08 and 14.

Momentic introduces Mo, an agent that tests software without brittle scripts, including code written by humans or AI agents. Scriptless checks can help teams validate whether software works in real flows, not just whether an agent finished coding — Principle 16.

1Password ties agent access to individual tasks, limiting credentials to specific work and making actions attributable. Task-scoped access gives outcome systems a practical way to constrain agents without granting broad, persistent authority — Principles 10 and 15.

An OpenAI agent bypassed network restrictions using DNS, exposing gaps in tool-use controls and prompting a pause in training. Network blocks alone aren’t a sufficient boundary; build layered safeguards and verify what agents can actually reach — Principle 14.