Salesforce adds tools to Agent Fabric that cap AI spending and stop failing agents
Salesforce added hard budget limits, agent testing and a kill switch that takes faulty agents offline to Agent Fabric, its AI control plane.
On October 27, 2026, Salesforce announced cost control and quality testing features for Agent Fabric, its platform for managing AI agents. The FinOps Dashboard, Model Wallet and Agent Kill Switch are generally available; Agent Evals enters beta in November 2026 and general availability in January 2027.
In an announcement on its newsroom on October 27, 2026, Salesforce said it had added new cost and quality features to Agent Fabric, the control plane it offers organizations for managing and governing AI agents across platforms. The company says Agent Fabric handles more than 1.2 billion requests a day. Salesforce also says that under half of organizations using agents test them on live traffic and that 96 percent of enterprises run agents on more than one platform.
- With the FinOps Dashboard and Model Wallet, budget owners can track AI spending on one screen and set hard caps so requests are refused before the budget runs out; both are generally available.
- According to Salesforce's test scenarios, Cost Optimization Policies can lower token consumption by as much as 75 percent through techniques such as semantic caching, payload compression and response trimming.
- Agent Evals tests agents, language models and MCP servers in production against a quality bar the team sets; if they fall below it, preset actions such as the Agent Kill Switch can kick in. Agent Evals enters beta in November 2026 and general availability in January 2027; the automatically triggered kill switch also becomes generally available in January 2027.
- Shared API policies can be applied at once across Azure, Apigee and Kong gateways, and a custom agent scanner, now in beta, adds agents behind firewalls to the catalog.
Why it matters
As AI agents go into use, model and token bills can become a cost line that is hard to predict. Salesforce's announcement may signal that controls such as budget limits, spending dashboards and automatic shutdown are becoming part of agent platforms. Companies in Türkiye starting agent projects could also benefit from setting a monthly spending cap, per-agent cost tracking and a procedure for taking an agent offline when quality drops, already during the pilot. The announcement does not include pricing.
For agent projects with costs and permissions defined from the start, see our custom AI agents for companies page.
Sources
More articles
Microsoft to process Claude models in Copilot inside the EU Data Boundary by the end of 2026
Microsoft says Anthropic's Claude models in Copilot will move into the EU Data Boundary. The target is the end of 2026, and the Claude setting is off by default in the EU and EFTA.
November 5, 2026 · Source: MicrosoftGoogle makes Gemini Omni 1.2 Flash video model generally available with clips up to 15 seconds
Google made its video generation and editing model Gemini Omni 1.2 Flash generally available on 5 November 2026. It is rolling out to the Gemini API, Google Cloud, Flow, Google Vids, the Gemini app and Google Ads.
November 5, 2026 · Source: Google BlogShopify's third-quarter revenue rises 32 percent to 3.74 billion dollars
In third-quarter results published on 3 November 2026, Shopify reported revenue up 32 percent and gross merchandise volume (GMV) up 29 percent. The company expects revenue growth in the low thirties percent again in the fourth quarter.
November 3, 2026 · Source: Shopify (SEC 8-K)