Weekly signal

This week (May 18–26, 2026) accelerated the practical rollout of agentic AI for real business automation: major vendors shipped agent runtime/connectivity features that let agents act across UIs, pipelines, and on‑prem data; platform vendors added MCP-style bridges and governance tooling; and enterprise vendors pushed no‑code agent builders and evaluation/testing for production automation. The net: agentic automation is moving from pilots and demos to operational primitives you can integrate into business processes today.

What changed

1. OpenAI and Dell announced a partnership to make Codex available for hybrid and on‑prem enterprise deployments — explicitly positioning coding/agent models to run near enterprise codebases and systems of record so agents can act with local context and data controls.

2. Microsoft marked key Copilot Studio capabilities (computer‑use agents that can operate UIs) as generally available in the Power Platform release wave, letting agents emulate human interactions on legacy web/desktop apps without brittle selectors. That reduces the time and integration cost to automate UI‑only systems.

3. Automation Anywhere released its 2026 Agentic Process Automation (APA) enhancements (including AAI Code, EnterpriseClaw, Context Intelligence Graph, and AI Evaluations) to coordinate agents, automations, and people with built‑in testing and runtime evaluation.

4. Zendesk unveiled the "Autonomous Service Workforce": Agent Builder, omnichannel AI agents, MCP support and outcome‑based pricing—aimed at replacing ticket‑deflection bots with outcome‑driven, governed service agents.

5. Integration/platform vendors shipped MCP server and agent‑friendly integration features (SnapLogic’s May 2026 release and similar MCP server moves across integration stacks), enabling pipelines and connectors to be exposed to agent runtimes as callable, governed tools.

What to do with it

• Prioritize architecture changes: treat MCP servers, agent connectors, and on‑prem model links as first‑class integration layers and plan to instrument them for identity, auditing, and reversible control.
• Pilot computer‑use agents on low‑risk, high‑volume legacy tasks; require human‑in‑the‑loop for exceptions and audit trails from day one.
• Use vendor evaluation tools (AI Evaluations, process simulation) or build lightweight test harnesses to measure correctness, tool usage, and objective drift before promoting agents to production.
• If you need data sovereignty/compliance, evaluate on‑prem options (Dell + Codex) and verify connectors, model locality, and SLAs with vendors.
• Start building a small catalog of MCP‑exposed tools (search, query, transaction) and define per‑tool permissions and toxicity/authorization guards before broader agent access.

Sources: see numbered list in the sources array below.