Build agents around proof, permissions, and parallel work

Anthropic redesigns Claude Projects for parallel work. Claude Projects turns one work description into coordinated parallel threads through Claude Code, making orchestration a first-class workflow rather than a prompt trick. This is Principle 09 in practice: define handoffs, context boundaries, and integration points before scaling agent volume.

The Feature Works. But Does It Belong?. CodeRabbit argues that agents can implement features faster than teams can decide whether those features serve the product, so evidence-based review and human approval remain essential. Outcome engineers need a Principle 15 gate that checks intent and user impact, not just whether tests pass.

UN System Data Commons Makes Global Statistics Accessible to AI Agents. The UN and Google make authoritative global statistics searchable through natural-language queries, giving agents structured access to a stronger source of truth. The move puts Principles 02 and 11 into infrastructure: connect agents to governed data and knowledge graphs instead of relying on recalled context.

I Vibed a Proof of Conway’s Conjecture. Dan Abramov uses AI to formalize a proof candidate in Lean, producing an artifact that a machine can check while independent mathematicians validate the result. For outcome engineers, Principles 02 and 16 point to the right standard: agent output should be executable, inspectable, and independently verifiable.

ZCode, the GLM Coding Agent, Silently Uploads Your Git History. An investigation reports that ZCode uploads encrypted workspaces containing complete Git histories, exposing the data boundary hidden inside a coding-agent workflow. Treat every connector and coding tool as part of the Principle 14 immune system: audit egress, restrict permissions, and establish trust boundaries before deployment.