Google PM open-sources Always On Memory Agent, ditching vector databases for LLM-driven persistent memory — Google releases an always-on LLM memory agent that stores structured memories without vector DBs. Outcome engineers can adopt a simpler, production-ready persistent-memory pattern that reduces infra complexity and improves long-lived agent state management (Principles 06 & 11).

Codex Security: now in research preview — OpenAI debuts Codex Security, an agent that finds vulnerabilities grounded in project-specific context and proposes safer fixes. This shifts vulnerability discovery into reproducible agent workflows that validate findings and cut triage noise, a practical step toward an engineering-grade immune system for agent-driven apps (Principles 14 & 16).

How Balyasny Asset Management built an AI research engine for investing — Balyasny ships a production AI research system that reasons like analysts and ties rigorous model evaluation to OpenAI-driven agent workflows. Use this as a template: embed evaluation, reasoning artifacts, and human-in-the-loop checks into agent pipelines to turn prototypes into auditable, tradeable outcomes (Principles 02, 03 & 09).

LangChain CEO: Better models alone won’t get AI agents to production — LangChain outlines Deep Agents that run long-lived tasks with isolated context, subagents, skills, and code execution. Outcome engineers must design the harness—context isolation, skill interfaces, and orchestration primitives—because model improvements alone don’t solve durability, safety, or observability (Principle 09).

Beyond ‘Prompt Thrash’: A Framework for Moving Agents from Demos to Production — An operational framework recommends treating agent quality as engineered via Outcome Specs and Convergence Loops to evaluate and prune configurations. Add explicit acceptance criteria, automated convergence checks, and short feedback loops to push agents from brittle demos into stable production (Principles 06, 14 & 16).