Salesforce 為 Agent Fabric 加入 Agent Evals、API Evals 及成本控制

Salesforce announced on October 27, 2026, a set of additions to Agent Fabric, its AI control plane for managing and governing agents across platforms, headlined by Agent Evals, API Evals, and advanced cost controls. Salesforce said the platform handles more than 1.2 billion requests daily.

The announcement frames the additions around a specific operational problem: as enterprises scale AI deployments, agents can fail silently, delivering wrong answers and accumulating costs before anyone notices. Salesforce said the new capabilities let CIOs and AI leaders catch problems before they turn into customer complaints or runaway costs, pairing automated interventions when quality slips with hard spending limits that cap usage at budget.

Agent Fabric is an enterprise agent control plane, built on MuleSoft, for discovering, governing, orchestrating, and observing agents, AI assets, and agentic workflows from a single control plane. According to the Agent Fabric product page, its scanners automatically detect and register agents, Model Context Protocol (MCP) servers, and APIs running across ecosystems such as Amazon, Google, Microsoft, Claude, Databricks, and Kong directly into Agent Registry, a centralized catalog of AI assets. The platform also includes MCP Bridge, which exposes existing APIs as MCP tools; multi-agent orchestration; support for the A2A protocol; a unified LLM and MCP gateway; Trusted Agent Identity; and Agent Visualizer, which maps the agent network with metrics including confidence scores, bottleneck detection, and hallucination risks. The product page states that Agent Fabric and Agentforce, Salesforce’s platform for building and running agents within Salesforce, are powered by the same engine for defining how agents handle interactions, reason, plan, and act.

Continuous Evaluation and Automated Intervention

Agent Evals automatically generates tests before deployment and then evaluates agents, large language models (LLMs), and MCP servers in production against defined quality thresholds, Salesforce said. When results fall short, the system triggers predefined interventions, including the Agent Kill Switch, with an approval step where a team’s policy requires one. The product page describes the capability as turning agent quality tests into automatic action, and lists MCP server and LLM evaluations as coming soon.

API Evals addresses a different failure point: agents break when the APIs they depend on change. Salesforce said the tool automatically generates functional and performance tests from API specifications, catching breaking changes before they reach production.

The announcement walks through a retailer scenario to illustrate the workflow. A retailer switches its order-support agent to a lower-cost AI model to reduce operating costs. After the change, Agent Evals runs scheduled tests and detects that the agent is providing inaccurate information about delayed orders. The system flags the quality issue and, based on the team’s preference to keep a human in the loop, requests approval to quarantine the agent. Once approved, the agent is quarantined while the team refines its instructions for handling delayed orders, reruns the evaluation on the lower-cost model to confirm performance, and returns the agent to production.

The Agent Kill Switch, which the product page describes as a way to quickly stop agent activity in production, is generally available as of the announcement. A Triggered-Agent Kill Switch follows to general availability in January 2027.

Cost Controls and Budget Enforcement

On the cost side, the FinOps Dashboard and Model Wallet give budget owners a single view of AI spending and hard limits that stop requests before costs exceed budget. Salesforce said Cost Optimization Policies cut token usage by up to 75%, based on the company’s test scenarios, through semantic caching, payload compression, and response trimming.

MuleSoft’s cost-management documentation details the enforcement mechanics. Model Wallet enforces budgets per agent, per instance, or per business unit, tracked live as spend accrues; when a wallet runs down, enforcement engages before an overrun rather than after the invoice, and forecasting flags a projected overrun days ahead. Model Proxy sits in the traffic path and prices each request against the rate the provider actually charges, so dashboard spending reconciles against the invoice. The optimization policies (Progressive Disclosure, Smart Response Trimming, and TOON compression) run at the gateway and cut token waste without modifying MCP servers or client code. The documentation also describes tracking spend by team, application, and use case in real time, and automatically shutting down runaway agent loops before they exceed the budget.

The FinOps Dashboard, Model Wallet, and Cost Optimization Policies are generally available as of the announcement.

Discovery and Federated Gateway Policies

Salesforce also announced two updates that extend Agent Fabric beyond its out-of-the-box platform support. Custom Agent Scanner brings automated discovery to internal environments, systems behind firewalls, and third-party platforms, letting teams build custom scanners that catalog agents, models, MCP servers, and APIs in Agent Registry without registering each asset manually. It is in beta as of the announcement, with general availability planned for January 2027.

Open Policies for Federated API Gateways discovers, enforces, and tracks API policies across gateways from Azure, Apigee, and Kong, applying a policy such as IP allowlisting or rate limiting across all of them in one action instead of one at a time. The capability is generally available as of the announcement.

Survey Data and Customer Statements

The announcement ties the additions to survey data on evaluation practices, noting that fewer than half of organizations with agents evaluate them on live traffic. It cites LangChain’s State of Agent Engineering report, which surveyed more than 1,300 professionals between November 18 and December 2, 2025. That report found 57.3% of respondents have agents running in production, 32% cited quality as a top barrier to production, and 89% have implemented some form of observability for their agents. On evaluation specifically, 52.4% of organizations reported running offline evaluations on test sets, while online evaluation adoption stood at 37.3%.

Poonam Deshpande, Director of Engineering, Digital Commerce at ASICS Digital, said ASICS is partnering with Salesforce to help shape the Agent Evals beta. “What excites us most about Agent Evals is that it doesn’t stop at telling us something is wrong,” she said. “When an agent’s quality drops, Agent Fabric can step in automatically, alerting the right owner or pulling the agent offline before a small issue becomes a widespread customer problem.”

The announcement also included statements from executives at Capita, GE Appliances, and Newell Brands. Tiina Stephens, Director of AI Engineering at Capita, said Agent Evals will help the company validate that agents perform against clear business and service outcomes. Adam Jones, Executive Director of Enterprise Transformation at GE Appliances, described a single control plane where the company can see its agents, who built them, and shut them down if something goes wrong. Steve Sigrist, Vice President of Global Customer Service at Newell Brands, said the company is using Agent Fabric as its control plane as it brings AI agents into customer service across its brands.

“Enterprises can only scale agents as fast as they can keep them accurate and on budget,” said Andrew Comstock, SVP and GM of MuleSoft and Agent Fabric. “Today, spotting a wrong answer or a spending spike is one job, and stopping it is another. Agent Fabric unifies them, bringing evaluations, cost controls, and intervention together so teams can contain issues instantly and scale agents with confidence.”

Availability

Agent Evals and API Evals enter beta in November 2026, with general availability planned for January 2027. Custom Agent Scanner is in beta as of the announcement and reaches general availability in January 2027, as does the Triggered-Agent Kill Switch. The Agent Kill Switch, Open Policies for Federated API Gateways, the FinOps Dashboard, Model Wallet, and Cost Optimization Policies are generally available as of October 27, 2026.