Salesforce switched Agentforce on by default. The toggle you lost is not the one that matters.

Salesforce is partway through rolling its Winter '27 release into production orgs, in staggered cutover weekends that started in the first week of September 2026 and continue into October. The change worth your attention: Agentforce now ships on by default in every org with Agentforce access, and the switch that used to gate it is gone from Setup.

Salesforce Ben reported the change in early September 2026. Every org on Enterprise, Performance, Unlimited or Developer Edition with Foundations, plus every Agentforce 1 org, gets the platform turned on automatically. The on and off toggle that used to live on the Agentforce Agents page in Setup has been removed.

That sounds more alarming than it is. Per Salesforce's own Winter '27 admin release notes, published in August 2026, turning the platform on does not turn an agent on. No agent starts handling a conversation or writing to a record purely because the org upgraded. Admins keep the Einstein Setup switch, the Manage AI Agents permission, and the decision to activate any specific agent. Salesforce built in a real control here, not a cosmetic one, and that is worth saying plainly before getting to what it does not fix.

This fits a pattern we have already written about. Agent platforms are shipping as infrastructure that exists by default rather than an add-on a team deliberately switches on, and the risk that creates is not that an agent does something malicious, it is that an audit trail built for named people quietly starts covering actions nobody consciously authorised. We covered the identity half of that problem in what changes about the system of record once agents start writing to it. Winter '27 is the access half of the same shift.

What actually changed on the way to production

The access decision moved from a page level toggle to a platform level default. That is a governance change, not an automation one.

Before Winter '27, an admin chose to expose Agentforce by flipping a switch on a specific page, and that choice doubled as a review moment: someone looked at the feature before it existed in the org. After Winter '27, the platform is present the moment the org upgrades, whether anyone looked at it or not. Salesforce Dictionary's coverage, dated 2 September 2026, independently confirms the same mechanics: the page level toggle is gone, platform access is default, and the rollout lands across several production weekends running into October 2026 rather than all at once. Salesforce Ben also reports the change carries no new licence cost and no billing change, which matters for the argument here: nothing about this is a commercial upsell forcing anyone's hand. It is a default, not a sale.

Why the admin checklists already out this month miss the point

A run of Salesforce blogs published an Agentforce Winter '27 checklist within days of each other. Nearly every one tells you to check the Einstein Setup switch. None of them ask who is allowed to activate an agent in the first place.

That is the wrong question to lead with. Our Revenue Orchestration Field Manual argues the decision that matters is not whether a platform is switched on, it is who holds the right to let an agent act, and what it is allowed to do once it can. A platform default does not grant an agent any new authority. It removes the one moment that used to force someone to notice the platform existed. It is the same pattern we described in why 2026 is the year internal apps become the control room: the page that grants access and the system that governs it stop being the same thing, and most teams keep checking the one that used to matter.

We made the underlying argument before Salesforce changed anything, in the rule for what you should never hand an agent: judge every task by reversibility and blast radius, not by whether the capability exists. An org that has never reviewed who holds the Manage AI Agents permission is one missed review away from finding out the hard way which of its agents can write where.

Picture the ordinary case rather than the dramatic one. An admin activates a lead qualification agent in a sandbox to pilot Agentforce ahead of Winter '27, the pilot gets judged a quiet success, and nobody revisits the activation once attention moves elsewhere. Before Winter '27, that agent still needed the page level toggle switched on to exist in the org at all, so a second person flipping that switch acted as an accidental check. After Winter '27 the platform is already present, so the only thing standing between a forgotten pilot agent and a live write is whether anyone remembered to look.

What to check this week, beyond the settings page

Three checks matter more than whether the toggle is on, because the toggle is gone either way.

• Who currently holds the Manage AI Agents permission, and whether that list was reviewed on purpose or inherited from an old profile.
• Which agents, if any, are already activated in the org, not just licensed, and what each one is allowed to write to.
• Whether each activated agent would pass a reversibility test before it touches the system of record, the same test behind the decision rights matrix in our B2B Revenue Automation Playbook.

None of that shows up on the Agentforce Agents page, because Salesforce already removed the one control that used to live there.

Winter '27 did not make Agentforce dangerous. It made the absence of a governance review invisible, by removing the page that used to force one. The checklist that matters is not the one that ships with the release notes.