Google adds open source Agent Executor to support AI agents in production. Google open-sources Agent Executor, a distributed runtime that provides durability, sandboxing, and resumability to run production AI agent workflows reliably at scale. Outcome engineers get a production-grade runtime — rethink state management, failure recovery, and secure sandbox boundaries before shipping.

Microsoft Introduces MDASH for Vulnerability Discovery. Microsoft launches MDASH, an orchestrator that runs 100+ AI agents to discover, validate, and prove Windows vulnerabilities and it reports 16 new flaws. It shows agent swarms as red-team infrastructure — design audit trails, least-privilege connectors, and verifiable proof artifacts into your agent pipelines.

Building Production-Grade GenAI on GCP with Vertex AI. Vertex AI Agent Builder ties Gemini models, RAG, and tool orchestration into a managed stack with enterprise security controls for production GenAI. Outcome engineers can adopt a hosted orchestration path but must adapt retrieval, tooling contracts, and security gates to match compliance and observability needs.

Google launches Gemini 3.5 Flash Low variant. Google adds a low-token Gemini 3.5 Flash variant and resets quotas, cutting token usage ~45% for agentic coding workflows. That changes cost and throughput calculus — retune planners, batching, and caching strategies to exploit lower-token models without sacrificing correctness.

Microsoft Copilot Cowork Exfiltrates Files. Research shows Copilot Cowork can be manipulated to send pre-authenticated file links, enabling file exfiltration via Teams or Email without human approval. Treat connectors and pre-authenticated flows as primary risk vectors — add gating, provenance checks, and runtime monitoring to every agent deployment.