Weekly signal

This week (July 20–28, 2026) delivered platform-level plumbing and model upgrades that materially change how enterprises build, run, and secure agentic business automation. Key developments: Microsoft Foundry pushed production-ready hosting, toolboxes and publishing paths into Microsoft 365 (making agents first-class automation endpoints inside Outlook/Teams/Office); Google’s Gemini Agent Platform hardened and advanced Memory Bank and event ingestion while patching an SSRF vector in Agent Studio and promoting new Gemini Flash models with built-in computer-use (browser/GUI) capabilities for automation tasks; Anthropic released Claude Opus 5—a more efficient, agentic model that vendors report improves end-to-end automation benchmarks like Zapier’s AutomationBench and OSWorld (computer-use); and Palo Alto Networks published Cortex AgentiX updates aimed at agent-driven SecOps automation and an “Agentic Assistant” feature for investigative workflows (July 24 release).

What changed

1. Platform-to-endpoint integration: Microsoft Foundry’s hosted agents, Toolboxes (skills/connectors), and routines reached production-grade maturity and now include a governed publishing path into Microsoft 365 Copilot and Teams — meaning agents can be deployed as governed automation endpoints inside the apps where work happens, with identity/permissions and lifecycle flows managed centrally.

2. Memory and event-driven agents: Google’s Gemini Agent Platform made Memory Bank profiles and the IngestEvents API broadly usable, offering low-latency structured memory and continuous event streaming for agents that must act on changing business state; the platform also shipped a security fix (SSRF) in Agent Studio on July 20. Gemini’s Flash-family models now include native Computer Use support for robust browser/GUI automation.

3. Model-level capability shift: Anthropic’s Opus 5 (July 24) narrows the gap between everyday models and frontier models for multi-step automation—customers cite better verification, fewer iterations, and superior computer-use performance for end-to-end business tasks.

4. Security & SecOps automation: Palo Alto’s Cortex AgentiX July 24 updates formally add agentic automation constructs (Cortex Agentic Assistant, automation toolchains) focused on investigation and response workflows.

What to do with it

• Evaluate integration, not just model accuracy: prioritize platforms that give agents safe publishing and identity (e.g., Foundry -> Microsoft 365) when planning automation pilots. Start design conversations with Microsoft tenant admins and security teams now.

• Instrument memory & event streams: for long-horizon automations (monitoring, churn automation, continuous compliance), test Memory Bank profiles + IngestEvents to reduce latency and query cost vs ad-hoc retrieval. Add end-to-end observability on ingestion windows and memory revision policies.

• Re-run your automation evals on Opus 5 and Gemini Flash (computer use): move at least 1 production-critical automation (e.g., billing reconciliation, customer onboarding flow, or automated regression testing) through a controlled A/B eval to measure pass rate, token/cost, and failure modes.

• Harden agent attack surface: treat computer-use and browser automation as new C2 vectors—apply allowlists, test for SSRF and similar injection paths, and bake agent-centric threat modeling into deployment gates. Regenerate Agent Studio web apps created before July 1 if you use Google’s Agent Studio.

• Security automation first steps: SecOps teams should evaluate Cortex AgentiX’s agentic assistant for orchestration of triage playbooks; pair it with centralized governance and human-in-the-loop approvals for high-risk actions.