Google Opens Gemma 4 Under Apache 2.0 with Multimodal and Agentic Capabilities — Google opens Gemma 4 under Apache 2.0, releasing multimodal, agentic open-weight models with up to 256K context and 31B parameters. Open, large-context agentic models let outcome engineers run, fine-tune, and orchestrate agent brains on-prem or in private clouds, reducing dependence on closed APIs and enabling new customization patterns (Principle 09).

AWS Launches Agent Registry in Preview to Govern AI Agent Sprawl Across Enterprises — AWS launches an Agent Registry preview to centrally discover, govern, and reuse enterprise AI agents, with native MCP and A2A protocol support. A centralized catalog and governance plane gives teams a practical way to control agent sprawl, enforce policies, and share vetted agent artifacts across orgs — core operational infrastructure for scaled agent orchestration (Principles 09 & 10).

Agents That Remember: Introducing Agent Memory — Cloudflare launches Agent Memory beta: a managed retrieval-based persistent memory that lets agents recall context without filling model windows. Persistent, retrieval-based memory is a foundational primitive for stable long-running agent behaviors, reproducible outcomes, and audit trails — a must-have for outcome engineering (Principle 11).

Android CLI: Build Android apps 3x faster using any agent — Google ships Android CLI and a Knowledge Base to speed agent-driven app development, cutting setup time and LLM token use while surfacing recommended, reproducible workflows. Opinionated, grounded developer tooling shows how to integrate agents into CI/CD and developer loops reliably — a blueprint for building agent-native developer experiences (Principle 07).

Anthropic Introduces Agent-Based Code Review for Claude Code — Anthropic launches agent-based Code Review for Claude Code, running parallel AI reviewers to find, verify, and prioritize pull-request issues in about 20 minutes. Multi-agent review pipelines demonstrate how automated verification, prioritization, and human-in-the-loop gates can be composed into delivery processes to raise velocity without losing safety (Principles 03 & 15).