How we made GitHub Copilot CLI more selective about delegation — GitHub tightened Copilot CLI’s delegation logic, reducing unnecessary subagent handoffs and parallelism that caused tool failures and latency. This matters because outcome engineers should prefer selective delegation patterns to cut failure modes and cost, an implementation of Agentic Coordination and frozen gates in practice (Principle 09).

NVIDIA Blackwell Leads on First Agentic AI Infrastructure Benchmark — NVIDIA’s Blackwell GB300 NVL72 posts up to 20x more concurrent agents per megawatt on the new AgentPerf benchmark. That changes capacity planning: you can realistically scale agent fleets at much lower power and cost, so update your deployment economics and runbooks for agentic infrastructure (Principle 07).

AI Agents Still Can’t Stop Prompt Injection Attacks, Researchers Warn — Researchers show agents built on GPT-5 and Gemini remain highly vulnerable to prompt-injection, with direct attacks succeeding over 79%. Outcome engineers must prioritize input sanitization, provenance, and active defense tests as part of an immune system for agents—these are operational must-haves, not optional audits (Principle 14).

architect-loop: Repo-centered Claude Fable planning with Codex builders — architect-loop wires Claude Fable planning to GPT-5.5 Codex builders using repo-driven specs, frozen gates, and sandboxed worktrees to cut token waste and enforce code review. Use this pattern to turn agent plans into auditable artifacts and reduce blast radius from automated changes, a practical Teamwork + Artifacts pattern (Principle 03).

Paca — Lightweight AI-native Jira alternative for human-AI collaboration — Paca treats agents as full Scrum teammates in a self-hosted project-management platform with sandboxed WASM plugins. If you’re designing human-agent workflows, study Paca’s integration and sandbox model for combining developer controls, auditability, and lightweight orchestration at team scale (Principle 03).