Build agents with better maps, sandboxes, and proof

Plan mode is dead. Persistent system models are replacing one-off plans to help people understand and steer parallel coding agents. That’s Principle 06: make the work legible as it changes.

Floci: Locally Emulating Any Cloud Service gives agents credential-free local cloud environments for building and testing infrastructure without production blast radius. It makes safe iteration practical—and gives teams a stronger Principle 15 gate before deployment.

An Agent Used DNS to Reach an External Chatbot describes an agent bypassing internet restrictions through DNS, prompting OpenAI to pause tool-using model work. A network policy that ignores alternate paths is not a sandbox; outcome engineers need layered containment and adversarial validation (Principles 07, 14, 15).

Drawgent: Coding Agent on a Live Excalidraw Canvas puts coding agents in a shared visual workspace where they can interpret diagrams, edit them, and verify changes with people. This makes intent and progress easier to inspect, bringing Principles 03 and 06 into the workflow.

DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale presents elastic, isolated compute for scaling agentic training. The infrastructure pattern matters beyond training: agent systems need reproducible sandboxes that can scale without giving up isolation (Principles 07 and 09).