Something shifted in the first two weeks of September, and it wasn’t a single announcement. It was the pattern.
On September 9, OpenAI launched a Data agent inside ChatGPT Work, letting businesses connect their own data sources, ask questions in plain language, and get back interactive dashboards. Two days later, OpenAI put its managed Agents API into public beta — a single service handling orchestration, context, and execution infrastructure, the plumbing that until now every team had to build themselves.
Meta moved the same week with Muse, a personal agent that searches the web, sends email, and makes purchases on a user’s behalf, shipping inside WhatsApp and as a standalone app. Salesforce introduced seven named Agentforce agents — Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin — each scoped to a specific job in sales, service, commerce, IT, HR, or supply chain.
Notice what’s absent from that list: capability claims. Nobody announced a smarter model. They announced org charts.
The interesting part is the governance
The Salesforce release included something easier to overlook than a cast of named bots: a control plane for registering agents, assigning identity and policy, tracking behavior, and controlling cost across both Salesforce and third-party AI.
That is not a feature anyone builds for a pilot program. You build it when you expect dozens of agents running at once and you need to answer who did what, under whose authority, with what audit trail. Industry forecasts now put roughly 40% of enterprise applications including task-specific agents by the end of this year, up from under 5% in 2025. Whether that number holds is anyone’s guess, but the infrastructure being shipped suggests vendors believe it.
What this means if you’re not a Fortune 500
The honest answer for most people reading this: not much yet, and that’s fine.
The useful move isn’t adopting an agent platform. It’s picking one repetitive task you already do — meeting notes, first-draft research, support triage — and testing whether a supervised AI version is actually better, with a human checking the output. Measure it against how you did it before. Keep what wins.
The failure mode this year isn’t falling behind. It’s handing a system broad permissions on a process you never bothered to define, and discovering the speedup was in making mistakes.
What are you actually using agents for? Bring it to the next meetup — we want to hear the unglamorous wins.
Leave a Reply