AI agents crossed from demo to deployment

What happened

Tom's Guide reported on 9 September 2026 that Meta launched Muse, a personal AI agent that can open websites, fill forms, send emails, book travel, and make purchases through Stripe's Link payment system using one-time-use card numbers. Muse runs inside an isolated cloud VM, a separate agent called Sentinel reviews actions before they reach the internet, sensitive actions ask for permission, and Meta publishes an audit trail of what the agent did and plans to do.

OpenAI introduced GPT-6 Astra in early September 2026, describing it as its strongest model for computer use and professional work. Two details matter to buyers more than benchmark tables: new enterprise admin controls (restricting approved sites and applications, controlling uploads and downloads), and confirmation policies that can require human approval before consequential actions. Enterprise access ships off by default.

A secondary report from 9 September 2026 says Databricks launched Agent Bricks, a platform for building domain-specific agents, and that more than 100,000 agents have been built on it since launch. Its KARL retrieval agent is claimed to match a frontier model's accuracy at lower cost and latency. These are vendor-supplied figures reported through a third-party article; treat them as unaudited until Databricks' own materials are checked.

The adoption numbers behind the moment

BetterCloud's June 2026 industry analysis aggregates the shift: Gartner expects 80% of enterprises to have deployed GenAI-enabled applications in 2026, Deloitte projects the agentic AI market growing around 53% annually from $8.5 billion in 2026 to $45 billion by 2030, and IDC forecasts more than 1 billion actively deployed AI agents by 2029.

• Trust asymmetry (per BetterCloud, citing its own sources): 84% of IT leaders say they trust AI agents as much as or more than humans for effective performance, while only 31% of employees are enthusiastic about them.
• The harder number: only 6% of companies fully trust agents to autonomously execute core business processes (reported by BetterCloud).
• Governance, data exposure, and shadow AI are named as the biggest barriers as agent adoption accelerates (BetterCloud).

What this means for a buying decision

The agent conversation has moved from "does it answer well?" to "what is it allowed to do?". That is a different checklist, and the vendors above have started writing it themselves: permission gates for sensitive actions, isolated execution environments, admin scoping controls, and visible audit trails.

Aviy's April 2026 trend guide (a vendor blog, so use it for framing rather than fact) captures the same caution we hold: fully autonomous business software is not here, and skepticism toward anyone promising it this year is warranted. Rule-based, repetitive work — reminders, follow-ups, recurring documents — is the credible first wave.

• Before evaluating an agent feature, list the actions it must never take unsupervised: payments, deletions, credential use, outbound communication.
• Ask for the audit trail before the demo. If a vendor cannot show what the agent did and why, the autonomy is not production-ready.
• Check the identity model: how does the agent hold credentials, and can it see them? Meta states Muse never sees underlying passwords or card numbers; ask every vendor the same question.
• Pilot on reversible tasks. The 6% full-trust figure is a reasonable mirror of your own risk posture.