Salesforce Agent Fabric Kill Switch Explained: Agent Evals, Hard Spending Caps and the January 2027 Trigger

on

salesforce agent fabric kill switch explained

Illustrative image: a data-centre server room. Agent Fabric is a software control layer, not a device, so there is no product photo for it — this image stands in for the infrastructure its agents run on. Image: Wikimedia Commons, CC BY-SA 3.0.

TL;DR: Salesforce Agent Fabric kill switch explained: Salesforce's cross-platform control layer for AI agents now pairs live evaluation with hard spending limits. The basic Agent Kill Switch is already generally available, Agent Evals enters beta in November 2026, and the automatic Triggered-Agent Kill Switch follows in January 2027.

📌 KEY HIGHLIGHTS

  • Salesforce announced the Agent Fabric expansion on 27 October 2026: Agent Evals, API Evals, a Custom Agent Scanner, Open Policies for Federated API Gateways, and advanced cost controls.
  • Agent Evals auto-generates tests before deployment, then keeps scoring agents, language models and MCP servers in production against quality thresholds you set.
  • When scores fall short, predefined interventions fire — up to and including the Agent Kill Switch, with human approval where your policy requires it.
  • Already generally available: the basic Agent Kill Switch, FinOps Dashboard, Model Wallet hard spending limits, and Cost Optimization Policies.
  • Agent Evals, API Evals and the automatic Triggered-Agent Kill Switch reach general availability in January 2027; the evals enter beta in November 2026.

What happened?

AI agents fail quietly. That is the problem Salesforce says Agent Fabric is built to fix. An agent can start giving wrong answers or burning through tokens long before a customer complains or the monthly bill arrives — and because agents now run across Salesforce, hyperscaler and home-grown platforms, finding the owner and getting permission to stop one takes time the whole time the bad answers keep going out.

Agent Fabric is Salesforce's MuleSoft control plane for discovering, governing, orchestrating and observing agents across platforms, and Salesforce says it already handles more than 1.2 billion requests a day. The 27 October 2026 expansion adds the missing loop: evaluation that acts, not just reports. Agent Evals generates tests before an agent ships, then keeps evaluating it in production. API Evals does the same for the APIs agents depend on, generating functional and performance tests from API specifications so a breaking API change is caught before it breaks the agent. A Custom Agent Scanner, now in beta, catalogues agents, models, MCP servers and APIs even in internal environments and third-party platforms Agent Fabric does not support out of the box.

Salesforce Agent Fabric kill switch explained: how the shutdown actually works

Here the details matter, because "kill switch" sounds more absolute than it is. The basic Agent Kill Switch became generally available in September 2026. It works through gateway controls that block governed traffic, and an intervention can be scoped to a request, a session, an agent or a tenant, with access restored afterwards. Blocking that traffic does not necessarily terminate every process running on the external platform hosting the agent — it stops the agent's governed activity, which is the practical off-switch most teams need.

What is new on the roadmap is the Triggered-Agent Kill Switch, due for general availability in January 2027. It connects evaluation findings to policy-driven intervention automatically: an agent that starts producing incorrect answers or consuming excessive resources can be quarantined without someone first noticing, finding the owner and filing a request. Salesforce's own example is a retailer that moves an order-support agent to a cheaper model to save money; scheduled tests spot inaccurate answers about delayed orders, Agent Evals asks for approval to quarantine the agent, the team fixes the instructions, and re-runs the tests before the agent returns to service. Human sign-off stays in the loop wherever a company's policy requires it.

salesforce agent fabric kill switch explained

Illustrative image: a person using an AI-assisted device. Image: Wikimedia Commons, CC BY 2.0.

Hard spending caps: the FinOps side

The second half of the announcement is about money. A FinOps Dashboard and Model Wallet give budget owners one view of AI spending, with hard limits that stop requests before costs pass the allocated budget — per agent, per instance or per business unit. Cost Optimization Policies use semantic caching, payload compression and response trimming, and Salesforce says they cut token usage by up to 75% in its test scenarios. Treat that 75% as a vendor test-scenario claim, not a production result: it applies only to traffic running through the configured policies, and customer numbers are not yet public.

Open Policies for Federated API Gateways, also generally available, lets a team push one policy — such as rate limiting — across gateways from Azure, Apigee and Kong in a single action, instead of configuring each gateway separately.

What is available now, and what is still coming

CapabilityStatus
Agent Kill Switch (basic, manual)Generally available since September 2026
FinOps Dashboard, Model Wallet, Cost Optimization PoliciesGenerally available now
Open Policies for Federated API GatewaysGenerally available now
Custom Agent ScannerBeta now; general availability January 2027
Agent Evals, API EvalsBeta in November 2026; general availability January 2027
Triggered-Agent Kill Switch (automatic)General availability January 2027

Why it matters for India readers

India is asking the same accountability question Agent Fabric tries to answer: who is responsible when an AI agent goes wrong? The Data Security Council of India has released industry guidance on liability for agentic AI, and the Telecom Regulatory Authority of India's draft guidelines on responsible AI in telecom — open for comments until 27 November 2026 — would require operators to keep an inventory of their AI systems, classify them by risk and report serious incidents, even when a vendor built the system. For Indian CIOs running agents across Salesforce, hyperscaler and home-grown stacks, a single inventory with live quality checks, audit trails and a tested way to switch an agent off is fast becoming a practical requirement rather than a nice-to-have. The timing also lands days before the US midterm elections, where AI chatbots were found giving wrong answers in 40% of election responses — a reminder that unchecked AI output is a live risk, not a theoretical one.

Should you care?

✅ Verdict: If your company runs AI agents in production, this is worth tracking — the generally available kill switch and spending caps are usable today, while the automatic evaluation loop is a January 2027 story. If you are not running agents yet, the lesson still applies: never deploy an agent without a way to test it live, cap its spend and switch it off. Salesforce has not published a standalone price for Agent Evals, so budget on the wider MuleSoft portfolio, not on a per-feature price tag.

FAQs

What is Salesforce Agent Fabric?
Agent Fabric is Salesforce's MuleSoft control plane for managing AI agents across platforms. It discovers, governs, orchestrates and observes agents, language models and MCP servers from one place, whether they were built on Salesforce or elsewhere, and Salesforce says it handles over 1.2 billion requests a day.

What is the Agent Kill Switch in Agent Fabric?
The Agent Kill Switch lets a team quickly stop an agent's governed activity in production through gateway controls, scoped to a request, session, agent or tenant, with access restored afterwards. The basic version is generally available; the automatic Triggered-Agent Kill Switch, which fires from evaluation results, arrives in January 2027.

What are Agent Evals?
Agent Evals automatically generates tests before deployment and then evaluates agents, large language models and MCP servers in production against quality thresholds the customer sets. When results fall short, it triggers predefined interventions, including an alert or the Agent Kill Switch, with approval where policy requires it.

When will Agent Evals be generally available?
Agent Evals and API Evals enter beta in November 2026, with general availability planned for January 2027, alongside the Custom Agent Scanner and the Triggered-Agent Kill Switch.

How do Agent Fabric's cost controls stop AI overspending?
The FinOps Dashboard and Model Wallet give budget owners one view of AI spending and set hard limits that block requests before a budget is exceeded. Cost Optimization Policies add semantic caching, payload compression and response trimming, which Salesforce says cut token use by up to 75% in its test scenarios.

TAKEAWAY: Agents fail silently and spend loudly. Salesforce's answer is a control layer that tests agents live, caps their budgets and can pull them offline — the manual version exists today, and the automatic version is the January 2027 date to watch.

Follow World Tech on TechByCountry for tomorrow's update.

📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.

Sources

Comments

Post a Comment