Business Automation Weekly AI News

September 14 - September 22, 2026

Weekly signal

Between Sep 14 and Sep 22, 2026 the business‑automation layer for agentic AI hardened into a stack that enterprises can operationalize. Major SaaS vendors shipped job‑ready agents and platform features built for long‑running business processes; RPA and low‑code vendors improved agent authoring, guardrails and voice/call handling; integration points let existing automations call agents directly; and a high‑visibility safety disclosure underscored the operational risk of testing agents with internet or tool access. The net effect: agentic systems are shifting from experimental assistants to composable automation components you can integrate, govern, and audit — but doing so requires changes to testing, orchestration, and security practices.

What changed

Salesforce (Sep 14) published a major Agentforce expansion: a portfolio of job‑ready agents (examples: Casey for service, Hunter for outbound sales, Marshall for supply chain, Paige for IT/HR) that ship with skills, actions and the data models required for those jobs and connect to Customer 360. Salesforce added a long‑horizon runtime so agents can pursue goals across days or weeks using memory, durable execution, and dynamic steering. The release also emphasizes Agent Script — an open, deterministic rule language to combine LLM reasoning with deterministic business rules — and multi‑agent orchestration to route work across specialist agents. The message: vendors want agents to take ownership of end‑to‑end business jobs, not just single prompts, and to keep auditability and business rules in the loop.

UiPath’s September notes show product movement that plugs agentic capabilities into traditional automation lifecycles. On Sep 14–16 UiPath announced conversational chat agents and real‑time voice agents within Maestro Flow, plus general availability for "Build agents with Coding Agents" — a lifecycle workflow that lets coding agents scaffold projects, edit prompts/schemas, run evals, package and deploy to Orchestrator. UiPath is shipping governance primitives too: Studio Web agent design policies (enforce min scores, temp ranges, escalation rules) and an "LLM as Judge" preview to use an LLM for content checks during evaluation. These features reduce the friction of moving agents from PoC to controlled production automation.

Microsoft’s Power Platform feature update (Sep 17) added explicit integration: you can now call Microsoft Copilot Studio agents from Power Automate flows and the platform shipped agentic authoring plugins and app‑building skills. Practically, that makes it easier to combine deterministic automations (flows) and long‑running agent work (Copilot agents) without custom middleware: agents become callable steps in existing automation orchestrations.

Anthropic published a Claude release note (Sep 15) that launched a Salesforce plugin (Salesforce in Claude, beta) with pre‑built sales skills that pull account/opportunity context into Claude. This is a cross‑vendor example of the same pattern Salesforce emphasized: connecting CRM data to agent skills so agents can act with business context rather than generic web knowledge.

Finally, a safety signal: Axios reported on Sep 18 that Google disclosed incidents from earlier testing where Gemini accessed three real companies during capture‑the‑flag tests because of environment misconfiguration. The disclosure is a reminder that giving agents tool or internet access without strong test isolation and controls creates real operational risk — especially when agents are used to act against enterprise systems.

Why this matters for automation teams

  1. Agents are becoming first‑class automation components. Pre‑built, job‑specific agents (Salesforce) plus callable agents in flows (Microsoft) and RPA‑native agent dev tooling (UiPath) mean teams can compose agent reasoning with deterministic automations and human approvals. That reduces integration work but raises expectations around governance and auditability.

  2. Long‑horizon work and durable execution change design patterns. Agents that maintain memory and run plans across days (Salesforce) need sensible lifecycle management: timeouts, retries, cost controls, and clear hand‑offs to humans and deterministic systems. Design patterns must cover persistent state and escalation.

  3. Authoring and CI workflows for agents are maturing. UiPath’s Coding Agents GA and guardrail tooling show the market moving toward a repeatable dev/test/deploy cycle for agents, not ad‑hoc prompt tinkering. Treat agents like software: versioned skills, eval suites, and deployment packaging.

  4. Security and testing are non‑optional. The Gemini disclosure shows the consequences of sloppy testing environments. Any agent with tool or web access should be tested in isolated environments and ramped up only after adversarial tests and environment hardening.

What to do with it (practical next steps)

  1. Prioritize two pilot scenarios: choose 1 high‑value long‑horizon business job (e.g., at‑risk deals rescue, multi‑touch outreach) and 1 high‑volume deterministic process (e.g., invoice triage, returns processing). Start with job‑ready agents (Salesforce Agentforce or vendor templates) where available to compress time to value.

  2. Add agent governance to your automation COE: require agent‑specific policies (min model score, allowed tools, max tokens, human approval gates), an LLM‑based content judge in CI for risky outputs, and session traces retained for audits. Use vendor guardrails (UiPath Studio Web policies, Copilot tenant controls) where possible.

  3. Integrate agents into existing orchestration: model agents as callable tasks in Power Automate/UiPath flows so deterministic actions remain traceable and agents handle reasoning. Design clear hand‑offs and durable state stores (memory, outcomes).

  4. Harden testing and production controls: run capture‑the‑flag style red‑team tests in isolated environments; validate that test data and test names do not map to real entities; disable internet/tool access until you have environment isolation and is‑the‑only‑one monitoring set up. Require post‑test remediation before enabling any external connectivity.

  5. Instrument and cost‑cap: enable fine‑grained telemetry (agent runs, tool calls, outcomes), set per‑agent cost caps, and track Agentic Work Units or equivalent metrics to assess ROI and runaway costs. Pair telemetry with session recordings and automated evaluators to spot false positives and silent failures.

  6. Plan for agent portability and skill reuse: where possible prefer open Agent Skill formats or Agent Script rules that can be ported or reimplemented across vendors to avoid lock‑in. When adopting a vendor agent, document which parts are proprietary (skill models, connectors) vs. portable rules.

Bottom line

This week consolidated the push to treat agents as production automation primitives: job‑ready agents, callable agent steps in flows, lifecycle tooling, and guardrails are now mainstream. That makes it faster to pilot and scale agentic automations — provided organizations adopt developer workflows, governance, isolation testing, and cost controls from day one.

Sources Salesforce — "Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High‑Value Work", September 14, 2026. UiPath — Agents release notes (September 2026): Maestro Flow chat/voice agents; Coding Agents GA; agent policies and LLM-as-Judge. Microsoft — Power Platform feature update, "What’s new in Power Platform: September 2026 feature update", September 17, 2026 (calls Copilot Studio agents from Power Automate). Anthropic / Claude — Release notes, "September 15, 2026: Launching Salesforce in Claude (beta)". Axios — "Google's AI hacked three companies in testing", reporting disclosure of Gemini incidents (Sep 18, 2026). UiPath — Product blog and product updates (context on Integration Service and automation orchestration for agents).

Weekly Highlights
Put an agent to work

Stop reading agent demos. Give one a job you repeat every week.

Describe the work, test the first result, and keep the agent available without running your own server.

Runs without your laptopBrowser + messaging appsManaged models includedMemory survives restarts

Plans from $130/month. Cancel anytime.

Hosted agent

OpenClaw or Hermes

saved state
Browser
WhatsApp
Telegram
Slack
“I checked the inbox, handled the routine messages, and sent you the one question that needs a decision.”
Create an AI worker that keeps running after this tab closes.
Open Agent Teams