Agent Optimizer: A Faster Path to Better Outcomes

Deploying an AI agent to production is not the finish line. Often, that’s when you start to understand what you’ve actually built. People ask for things you didn’t anticipate. A workflow that looked good in testing takes too long when it’s running live. Model behavior shifts, business requirements change, and something that worked last week needs attention this week.

At Engine, a travel management platform, agents resolve 50% of chat inquiries, with 15% lower customer support handle time. Hibbett, a sporting goods retailer, uses agents to handle 90% of its core shopper journeys. What we’ve learned from customers like these is that continued success depends on a fast iteration loop. Improving those outcomes means learning from what’s happening and putting those lessons back into the agent.

But for many customers, executing that loop today takes too much manual work. You read sessions, inspect failures and look for patterns. Then you update instructions or build a new action, test the change and check that fixing one problem didn’t create another. You also need to turn what you learned into regression tests so you can make the next change with more confidence. Meanwhile, your agent is handling thousands more conversations.

We believe the future is agents executing that loop. Developers are already beginning to feel that acceleration with coding agents like Claude Code and Codex. Bringing that success to every enterprise means pairing the power of agents with a UI that makes their work visible and gives people control. Users need to understand what the agent is doing, review its changes and choose how much it can do on its own. That’s how we build trust and make this capability accessible to less technical users, too.

People define the outcomes that matter, guide the work and decide which changes need their review. Agents investigate what’s happening, build improvements and test them.

That’s why we’re introducing Agent Optimizer, our agent for building and improving agents. It connects production insights with building and testing so teams can move from finding a problem to validating an improvement in one continuous workflow. It’s a step toward agents that learn from experience, with humans guiding that learning. Here’s how it works.

Build, test, and iterate in one conversational loop

Building an agent shouldn’t require you to learn every detail of Agentforce and Salesforce first. It should start with expressing your business needs in plain language: “Help me build an agent that can handle order returns.” With Agent Optimizer in Agentforce Builder, you express the business intent, and Optimizer translates it into agent configuration, including the subagents and actions needed to support it. You can refine the agent’s behavior through conversation, without having to architect the whole system yourself.

Critically, you don’t have to start with a blank page. Your existing customer conversations contain a lot of information about what your agent needs to handle. Optimizer can look at the reasons customers contact you, using transcripts of past conversations with agents or human support reps, and use those patterns to help build out the capabilities your business needs. That gives you a starting point grounded in what customers are actually asking for.

As Optimizer builds, it tests. It simulates conversations the same way you would manually preview an agent: make a request, see how the agent responds and work through what happens next. When something fails, it can investigate, make a change and try again. You can review the configuration alongside the conversations that show how it behaves.

Those previews and tests can become reusable regression tests. As you build out the agent, you’re also building the test suite that helps you improve it with confidence. The next time you change an instruction or add a capability, Optimizer can check that earlier test cases still pass. Building, testing and fixing become part of the same workflow, so you can get from a business need to a validated capability faster.

Turn production insights into better outcomes

The real work begins once the agent is live, with new insights, opportunities and issues emerging every day. You might see more escalations, lower customer satisfaction or customers abandoning conversations, without an obvious explanation for what’s going wrong.

In Agentforce Observability, Optimizer can analyze hundreds or thousands of sessions, identify recurring failure patterns and help prioritize the improvements likely to have the greatest impact. You define the outcomes that matter to your business, like resolution rate or CSAT, and use those goals to guide the investigation.

From there, you can work through the findings in conversation. Ask why customers are getting stuck and Optimizer will help you drill into relevant sessions and investigate the causes behind those patterns. Optimizer can group failures by tool-call errors, knowledge gaps or off-topic requests. It might uncover a knowledge action pointing to a data source the agent can’t access, or a recurring customer question your content doesn’t cover. You can inspect the evidence, understand the recommendation and decide what to address.

That investigation connects back to the building and testing workflow. Optimizer can recommend an improvement and hand it off for implementation and testing in Builder. What you learn from production becomes the starting point for the next iteration.

You choose how much of that loop happens autonomously. An “autonomy dial” lets you require sign-off at every step or give Optimizer room to investigate issues, suggest fixes, build, test and stage changes for deployment, stopping at the review points you’ve chosen. You can watch it work in the browser or step away and come back to review the results. You define what better looks like and how much control to delegate.

Toward self-learning, self-improving agents

Agent Optimizer is a step toward the future we see for the agent development lifecycle: self-learning, self-improving agents that learn from production, test potential improvements and build on what works. A continuous hill climb toward better business outcomes, with humans defining success and guiding the work.

The ambition is for every iteration to make the next one more effective. What an agent learns today should translate into better business outcomes tomorrow.

Agent Optimizer is available now in beta. Contact your account executive to learn more and check out the demo below.