Skip to content
Signals · Product

Salesforce's New Agent Just Debugs Other Agents

Auto-generated note

Salesforce has introduced Agent Optimizer, an AI agent purpose-built to find and fix failures in other production AI agents, connecting production insights directly to building and testing so teams move from spotting a problem to validating a fix in one workflow instead of juggling separate tools. A rep can describe a need in natural language — “help me build an agent that can handle order returns” — and Optimizer translates that into agent configurations grounded in historical customer conversations, plus reusable regression tests for iterating with confidence. Inside Agentforce Observability, it scans hundreds or thousands of production sessions, groups recurring failures by root cause (tool-call errors, knowledge gaps, off-topic requests), and prioritizes fixes against business metrics like resolution rate or CSAT. An autonomy dial lets teams choose oversight level, from mandatory sign-off to fully autonomous investigate-fix-test-stage cycles. Salesforce cites travel platform Engine (agents resolving 50% of chat inquiries, 15% lower handle time) and retailer Hibbett (agents handling 90% of core shopper journeys) as proof points. Agent Optimizer is in beta now, reachable through a Salesforce account executive.

The release lands the same month Salesforce disclosed just how much live agent traffic it’s already carrying. Casey, its Help Agent for small business service teams, has handled more than 4 million customer inquiries on help.salesforce.com alone at roughly a 70% resolution rate, running on pay-per-resolution pricing with no charge for cases escalated to a human.

Put together, that’s the actual news: Salesforce isn’t just adding more agents to Agentforce, it’s shipping the tooling to manage the ones already in production at Casey-level volume. For sales orgs building or scaling their own agent deployments, Agent Optimizer’s existence is itself a data point — once agent traffic gets big enough, debugging and observability stop being optional and become their own product category.