Salesforce Adds Agent Spending Limits, With Evaluation-Triggered Controls Still on Its Roadmap

Salesforce says new spending controls for Agent Fabric can stop AI requests before they exceed a configured budget. But the evaluation-driven workflow that detects declining agent quality and triggers an intervention is not generally available yet.

The distinction sits inside the company’s October 27, 2026 announcement: some cost and shutdown controls are listed as available, while Agent Evals, API Evals and the Triggered-Agent Kill Switch have later release targets. Buyers should not read the whole announcement as one finished product launch.

What Salesforce says is available now

In its dated availability schedule, Salesforce lists the FinOps Dashboard, Model Wallet, Cost Optimization Policies and the existing Agent Kill Switch as generally available. The evaluation products and triggered shutdown capability have a different timetable.

These dates are Salesforce’s announced schedule, not a hands-on confirmation of access in every customer environment.

A documentation mismatch buyers should notice

The MuleSoft cost-management page describes Model Wallet budgets scoped to an agent, instance or business unit. It says spending is tracked as it accrues and enforcement acts before a budget overrun.

However, the same page’s availability FAQ still says Model Wallet and the Unified Dashboard arrive later in the quarter. That does not match the October 27 announcement’s general-availability statement. We cannot establish from these pages alone whether the FAQ is outdated or reflects another rollout distinction. Customers should confirm entitlement and deployment requirements before relying on the control.

The scope also matters. This is a claim about governing AI usage costs, not a newly announced payment-card limit on what an agent can purchase from merchants.

Stopping spend is different from stopping poor work

Unite.AI’s October 27 coverage also separates the existing kill switch from its future triggered version. It describes Agent Evals as testing agent behavior against quality thresholds and initiating a predefined response, with approval where an organization requires it. API Evals is intended to generate functional and performance checks from API specifications.

Our analysis: A budget threshold and a quality threshold answer different questions. The first asks whether more usage is authorized. The second asks whether the agent is doing acceptable work. An agent can stay under budget while giving bad answers, or perform correctly while consuming more resources than the owner intended.

Likewise, having a stop control does not establish that failed evaluations automatically activate it today. That separation is relevant to the shutdown questions in our DeepSeek agent evaluation coverage and the identity-versus-containment distinction in IBM’s agent-control previews. These are different products and tests, not evidence that Salesforce’s implementation behaves the same way.

The savings claim needs an outcome check

Salesforce claims up to 75% lower token usage from optimization policies in its test scenarios. That is not a guaranteed reduction in a customer’s total bill or proof that task quality remains unchanged.

Our analysis: Fewer tokens are useful only if the workflow still delivers the required result. As our Nvidia–Meta infrastructure coverage explains, cost per successful, authorized task is a more useful question than efficiency considered alone.

What we would test before trusting the limits

In an authorized, low-budget test environment, we would check whether new requests stop at the configured threshold, how concurrent and already-running requests are handled, and whether retries or another route escape the intended budget scope. We would also compare recorded usage with provider charges.

When the evaluation-triggered controls become available, a separate test should introduce a known quality failure and measure detection, approval and enforcement time. It should also check the recovery process before the agent resumes. These are proposed tests; none was performed for this article.

For now, the release offers a clearer announced budget-control layer. The promised connection between quality checks and automated intervention remains a roadmap commitment, and the conflicting cost FAQ makes deployment confirmation especially important.