Weekly signal

This week (covering Sep 14–22, 2026) accelerated the push to make agentic AI a production automation fabric: large SaaS platforms shipped job-ready agents that can run multi-step business processes, RPA vendors tightened agent development and runtime tooling, platform-level connectors made it easier to call agents from existing automations, and safety disclosures reminded teams to bake adversarial testing and guardrails into deployments.

What changed

1. Salesforce launched a new Agentforce portfolio of job-ready, long-horizon agents for sales, service, commerce, back office and HR — including pre-built agents (Casey, Hunter, Marshall, Paige) that connect to Customer 360 and run across channels and workflows. Agentforce adds a long‑horizon runtime (memory, durable execution, dynamic steering) and Agent Script (an open language to combine LLM reasoning with deterministic rules).

2. UiPath released September agent updates: Maestro Flow now supports conversational chat agents and real‑time voice agents, and Coding Agents (scaffolding, evaluation, packaging and deploy flows driven by coding agents) reached general availability. Studio Web agent design policies and an “LLM as Judge” guardrail are also shipping or in preview to enforce agent configuration and evaluation.

3. Microsoft Power Platform’s September feature update added the ability to call Microsoft Copilot Studio agents from Power Automate, plus agentic authoring plugins and skills that enable AI-driven app building and process orchestration inside existing automation flows. This lowers the friction to orchestrate agents alongside Power Automate flows.

4. Anthropic launched a Salesforce plugin (Salesforce in Claude, beta) that brings CRM records and pre-built sales skills into Claude, enabling Claude-managed agents to act on sales context — a clear example of cross-vendor agent integrations for enterprise automation.

5. Google disclosed safety incidents in which Gemini accessed real companies during third‑party capture‑the‑flag testing; the disclosure (Sep 18) highlights operational risk from testing and environment misconfiguration when agents are given internet or external access.

What to do with it

• Map 1–3 priority use cases (long‑horizon work, buyer/seller outreach, order-to-cash) and test with a job-ready agent first before building bespoke agents.
• Add agent-level governance: enforce policy (min score, temperature, max tokens), human‑in‑the‑loop gates, audit logs and an LLM-as-judge evaluation step during CI for agents.
• Integrate agents with existing automation orchestrators (Power Automate, UiPath Orchestrator) to combine deterministic RPA with agent reasoning and durable execution.
• Run adversarial and environment‑isolation tests (capture‑the‑flag style) on any agent with tool or internet access; treat test configuration as a security-control checklist.
• Pilot with strong telemetry, cost caps, and a small set of enterprise data connectors; iterate on skills and Agent Script rules for predictable decisioning.