Agentforce in Production: How to Monitor, Test, and Improve AI Agents

Launching an AI agent in Salesforce is not the end of the project. It is the point when the most valuable information starts to appear: how users interact with the agent, where they get stuck, which requests trigger escalation, and which workflows actually perform well in real conditions.

That shift in focus was especially visible at Dreamforce 2026. Salesforce introduced capabilities aimed at operating production agents, including Agent Optimizer, Agent Health Monitoring, Custom Scorers, conversation testing, and A/B experimentation. The message is straightforward: building an agent is only part of the job; keeping it reliable and useful requires continuous observation and improvement.

After go-live, the problems change

During development, teams can control test scenarios. In production, users ask unexpected questions, data changes, new processes appear, and an instruction that worked well can start producing weaker outcomes.

Salesforce describes this as an iterative learning loop: observe production behavior, identify recurring failure patterns, test improvements, and turn lessons learned into regression tests. That loop is what separates an agent that was simply launched from one that continues to create value over time.

What should be monitored in production

• Error rate: helps identify interactions where the agent failed to complete an action correctly.

• Latency: shows how long the agent takes to respond.

• Escalation rate: indicates how often the conversation needs to move to a human.

• Response rate: helps detect cases where a user does not receive a response within the expected window.

• Resolution and quality: should be assessed against the actual business objective, not only the absence of technical errors.

Agent Health Monitoring is designed for this operational layer. Salesforce describes it as a way to detect issues such as higher error rates, latency spikes, or changes in escalation and response behavior, with configurable alerts and session-level investigation.

Agent Optimizer: turning real sessions into improvements

Agent Optimizer goes a step further. Salesforce announced general availability for October 2026 and positions it as an agent that helps teams build, test, and improve other agents.

Within Agentforce Observability, Optimizer can review large volumes of production sessions, group recurring failure patterns, and help prioritize the issues with the greatest potential impact. Salesforce examples include tool-call errors, knowledge gaps, and off-topic requests.

The practical difference is the connection between observation and execution. After identifying an issue, teams can work on the change, test it, and stage it within the same improvement loop, reducing the risk of fixing one problem while introducing another.

Testing still matters before every change

Automating parts of the improvement process does not mean giving up control. Salesforce continues to frame human oversight as central: teams define the outcomes that matter, decide how much autonomy to allow, and choose where review is required.

Other Dreamforce 2026 announcements reinforce the same idea. Custom Scorers let organizations evaluate sessions against their own criteria. Voice and Conversation Testing evaluates full conversations rather than isolated turns. A/B Experimentation, announced in beta for October, lets teams compare agent variants across traffic cohorts.

What this means for companies already using Agentforce

• Define the business outcomes the agent should improve.

• Set technical and operational metrics before scaling volume.

• Review sessions and failure patterns on a defined cadence.

• Test changes before production deployment.

• Keep data, knowledge, permissions, and actions aligned with the current business.

Where maintenance, support, and audit fit

This approach follows the same principle behind post-go-live Salesforce support: the platform should not stand still after implementation. With AI agents, that becomes even more important because quality depends on data, processes, permissions, and behaviors that evolve over time.

Azimute's maintenance and support services help the CRM evolve with the business, while technical and strategic audits identify critical issues, underused capabilities, and improvement opportunities. In an Agentforce environment, that discipline also means understanding whether the agent is responding, deciding, and acting in line with the intended outcomes.

An agent that learns needs an operation that learns too

Salesforce's new tooling can accelerate the improvement loop, but it does not remove the need for strategy. Technology can identify patterns and test hypotheses; the business still needs to define success, acceptable risk, and the points where human judgment remains essential.

If you already run Agentforce in production, or are preparing your first use case, the useful question is no longer only 'does the agent work?'. It is 'do we have the metrics and processes to know whether it is getting better?'.

Azimute can help assess the current state of your Salesforce environment, structure post-launch operations, and define an evolution roadmap aligned with your business.