Exploring and Setting Up Agentforce Observability
For Salesforce admins managing AI agents, the job does not stop at deployment. Once an agent is live, the real work begins: monitoring whether it is actually resolving issues, tracking where it falls short, and catching problems before they reach customers. That means keeping an eye on metrics like deflection rates, escalation frequency, CSAT scores, and task success rates: the numbers that tell you whether your agent is an asset or a liability. Until recently, getting that kind of visibility required piecing together data from multiple places.
Ever since its release in 2024, Agentforce and its affiliated AI tools reshaped how Salesforce works, transforming the core mechanism of the platform from declarative mechanisms and coding to prompts and agentic workflows. Despite all these capabilities, Agentforce suffered from one main weakness: its black box-like structure. Organizations could deploy agents and observe their outputs, but had little to no visibility into the reasoning and decision-making that produced those results… until the release of Agentforce Observability.
Salesforce responded to one of its most consistent pieces of user feedback by building a dedicated observability layer directly into Agentforce Studio, giving teams the visibility they had been asking for.
What Can Agent Observability Track? Core Metrics and Data Points
When an AI agent goes sideways, it rarely throws a clean error. It might return a confident but wrong answer, get stuck in a loop, or take an action no one anticipated, all while appearing to function normally. Agent observability exists to solve that problem. It exposes the decision-making process behind every action so teams can understand not just what an agent did, but why. Here’s what falls within its scope:
- Tool selection and execution: which tools the agent chose to invoke, when it called them, and whether those calls produced useful results or stalled the workflow
- Internal reasoning steps: the model’s decision process before it commits to an action, which is where silent failures like logically flawed but fluent responses tend to hide
- Prompt version history: a record of which system prompt was in use at the time of any given interaction, making it possible to connect behavior shifts to specific configuration changes
- Model parameter settings: values like temperature that shape how the model generates responses, captured per session so results can be reproduced and compared
- Grounding context: the specific data retrieved during an interaction, which surfaces whether the agent had accurate, relevant information to work from
- Latency and error signals: performance baselines that can indicate problems with connected APIs, infrastructure bottlenecks, or system resource issues
- Reasoning loop outcomes: whether the agent reached a resolution, repeated itself without progress, or called a tool that doesn’t exist
- Evaluation scores: human or model-generated grades on agent outputs that feed into an ongoing improvement cycle
Together, these data points give teams the full picture of what an agent is doing, how it got there, and where to focus when something needs to change.
Agentforce Becomes More Transparent
Agentforce Observability is your centralized application for fine-tuning your agents. It provides a single mission control for IT and business teams to get the complete picture of agent performance and impact. This includes deep observability, live health monitoring, rich adoption analytics, and consumption tracking; as well as capabilities to embed agent signals in the flow of work for line-of-business users like service leaders.
Follow the Step-by-Step Procedure Below to Access Agentforce Observability:
- Open the App Launcher.
- Search for or select “Agentforce Studio” from the App Launcher menu.
- Once inside Agentforce Studio, navigate to the “Observe” pane in the left-hand panel. This is where all Observability tools are housed.
- From the Observe panel, select the tool you need: Agent Analytics, Agent Optimization, or Agent Health Monitoring.
Salesforce has Organized Agentforce Observability into Three Distinct Functional Areas
Agent Analytics: provides a comprehensive view of how every agent is performing in real customer interactions and surfaces KPI trends over time.
Agent Optimization: records the steps in the agent’s reasoning chain so teams can understand why and how a particular agent behaved in a certain way. It observes every user interaction and has the capabilities to compile similar requests to identify patterns.
Agent Health Monitoring: which provides more transparency into agent health, allowing teams to visualize performance trends in real time and get alerted to failures when they happen.
Scorers: a beta function as of May 2026 when this article is written, can also appear in your left side pane. It is not strictly a part of Agentforce Observability. Since Testing Center has been integrated directly into Agentforce Studio as a dedicated tab alongside Agent Builder and Observability, Scorers may also appear alongside your usual Observability capabilities.
Setting it up
Unfortunately, Agentforce Observability is not an out-of-the-box feature of core Salesforce. It requires Data 360 (aka Data Cloud) to function. It also requires at least one agent of which it will track the statistics.
Ensure that you have Data 360 and Einstein enabled before starting the process. If you have not enabled Data 360, follow the steps below:
- Go to Setup → Data Cloud Setup (or navigate to it via the App Launcher)
- Click Get Started on the guided setup page to begin manual provisioning
- Wait a few minutes for the instance to initialize. This will enable Data 360.
- Go to Setup → Einstein Setup
- Click the Turn on Einstein toggle. This will enable Einstein.
Check your Salesforce Standard Data Model
Once you complete the steps above, make sure your org is running version 1.130 or higher. Navigate to Setup, then go to Setup → Apps → Packaging → Installed Packages. Then, look for the Salesforce Standard Data Model to verify the version number. If an upgrade is needed, Salesforce provides an installation link to bring it up to the required version.
The installation of Data 360 will give you several new permission sets. To complete the rest of the process, you need:
- to be a System Administrator Data Cloud User
- and Data Cloud Architect permission sets
To enable Agentforce Session Tracing, go to Setup → Einstein Audit, Analytics, and Monitoring Setup. Then, click on the Agentforce Session Tracing toggle. Since we will be using a newer version of Agent Analytics, clicking on the Agent Analytics toggle above is not required. Make sure to also enable Agent Optimization from here.
Go to Setup → Agent Analytics. You will see two list items: Employee Agent Analytics and Service Agent Analytics. Open their respective menus and click Install on both of them.
This will finalize the configuration of Agent Analytics for your org. The Analytics menu will now be enabled and show real data for your agent(s). You can filter the analytics shown by the agent of your choice, and also monitor how it has been used in unique sessions by unique users. More information about Agent Analytics can be found here or under the Agent Analytics Help button shown below.
Ready to See What Your Agents Are Really Doing?
Agentforce Observability marks a turning point in how organizations deploy and manage AI agents on the Salesforce platform. What was once an opaque system now has a dedicated suite of tools that bring transparency to every interaction. As Agentforce continues to evolve with features like Scorers already in beta, the Observability suite will only grow. If you run an organization that is serious about scaling AI responsibly, Agentforce Observability is your friend.
Explore related content:
How to Deploy Agentforce Right the First Time, Straight from the Experts
Choices, Choices: What Are Radio Button Groups Best Used For
