Verifiable Evidence in Agentforce Shows Exactly How Agents Reason and Act
Agentforce

Verifiable Evidence in Agentforce Shows Exactly How Agents Reason and Act

By Troy Amyett•November 13, 2026•9 min read
Speed
AI-narrated

Agentforce gives you concrete, inspectable records of how every agent reaches its conclusions and executes actions. The platform logs reasoning steps, selected tools, variable changes, source citations, and final outcomes in structured formats that admins and developers can review directly.

This approach replaces black-box outputs with traceable evidence across the full agent lifecycle.

What makes Agentforce evidence verifiable?

The system combines probabilistic outputs from large language models with deterministic controls such as scripts, filters, and explicit verification steps. Every interaction produces a session trace stored in a unified data model inside Data 360 (formerly Data Cloud). Platform traces add service-level details from Apex, Flows, Prompt Builder, and the LLM Gateway. Together these create an unbroken chain from user input to final result.

You can open any trace and see the exact prompt sent to the model, the variables before and after each action, the subagents invoked, and any errors encountered. The Einstein Trust Layer adds zero-retention policies for third-party models, dynamic grounding, and logs that record which data sources were used. This combination lets you validate accuracy, trace compliance issues, and identify where an agent deviated from expected behavior.

Consider a support agent that pulls account details before suggesting a refund. The trace shows the initial user message, the knowledge article retrieved for policy language, the exact variable that stored the account ID, and the point where the refund action received approval. If the agent later uses an outdated policy version, the same trace shows the timestamp mismatch between the article and the grounding step.

Traces also surface when an agent calls an external service. The log records the request payload, the response code, and the time taken. If the service returns an error, the trace captures the error message and the agent’s next decision, such as retrying or escalating. This level of detail supports compliance checks that require proof of which data source informed a particular recommendation.

How do traces capture agent reasoning step by step?

Session tracing records every turn in a conversation as a structured timeline. The Interaction Summary panel displays decisions in sequence, including which action or subagent was chosen, the full JSON of prompts and responses, and timing for each step. You can expand any node to inspect variable state changes or drill into error details without leaving the Agentforce Builder interface.

Platform tracing extends this view to underlying services. When an agent calls an Apex class or triggers a Flow, those executions appear with their own telemetry. The result is a single navigable record that shows both the high-level reasoning path and the low-level platform activity that supported it. In practice this means you can reconstruct exactly why an agent escalated a case or pulled information from a particular knowledge article.

Take an order-status agent that receives a customer inquiry about a delayed shipment. The first turn shows the agent selecting a knowledge lookup action. The trace then records the search term passed to the knowledge base, the top three articles returned with their relevance scores, and the article the agent ultimately cited. Later turns show the variable that held the order number and the moment the agent decided to create a case instead of answering directly.

The same trace also logs timing data. If a subagent takes eight seconds to respond, the record flags that latency and links it to the specific service call. Over multiple sessions you can spot patterns, such as repeated delays when the agent queries a particular external system. These patterns help you decide whether to adjust the action sequence or add caching logic.

Traces grow quickly in complex conversations. A single session with five turns and two subagent calls can produce several hundred lines of structured data. Reviewing every trace manually becomes impractical, which is why the analytics layer later aggregates the same records into summaries.

Does the Testing Center catch problems before they reach users?

The Testing Center runs batch evaluations against AI-generated or custom utterances before deployment. Scorers automatically assess response accuracy, completeness, coherence, conciseness, latency, and adherence to instructions. You can also define custom assertions that check whether specific subagents or actions were invoked correctly.

Agentforce Assistant (Beta) can generate additional test cases and suggest iterations based on scorer results. This process surfaces issues such as missing knowledge coverage or inconsistent instruction following while the agent is still in a sandbox. Teams that run these evaluations report clearer acceptance criteria and fewer surprises after go-live because the scorers quantify quality rather than relying on manual spot checks.

A practical test might include fifty utterances that ask the same policy question in different phrasings. The scorer flags any response that omits a required disclaimer or cites the wrong article. Custom assertions can verify that a verification subagent runs before any refund action executes. If three of the fifty utterances fail the assertion, the test report lists the exact prompts and the point where the agent skipped the verification step.

The Testing Center does not replace human review entirely. Scorers measure surface qualities such as coherence but cannot judge whether a response aligns with unwritten business judgment. Running the same test set after each prompt change shows whether the adjustment improved or degraded the scores, giving a concrete signal for iteration.

When should you add citations and verification actions?

Citations link every generated response to the specific knowledge articles, PDFs, or web pages used for grounding. You enable them either through platform-managed settings or explicit instructions in the prompt template. The links appear to users and remain available in the trace for later review.

Customer verification uses standard subagents that send a one-time passcode and confirm identity before allowing access to private data or actions. Once verified, the agent sets variables that downstream filters can check. These steps add measurable security controls without requiring custom code for every sensitive workflow. Use them when agents handle regulated data or when regulatory requirements demand proof of identity before any change occurs.

In a financial-services scenario the agent first confirms identity, then retrieves account balances only after the verification variable is set to true. The trace records the passcode sent, the response received, and the exact moment the variable changed. If an auditor later questions how a balance was disclosed, the trace supplies the verification timestamp and the source article that contained the balance data.

Citations and verification add steps to the conversation. Each extra turn increases total handling time and token usage. For agents that only surface public product information, the added overhead may outweigh the benefit. The decision hinges on the sensitivity of the data and the regulatory environment rather than a blanket rule.

What do analytics dashboards actually show after deployment?

The Command Center and Agent Analytics dashboards track outcomes such as resolution rate, escalation frequency, quality scores, adoption, and cost metrics. Real-time alerts notify you when latency or error rates exceed thresholds. The Optimizer (Beta) surfaces recommendations based on patterns across sessions.

Data flows into Data 360 and can connect to Tableau Next or external tools through OpenTelemetry. This setup supports root-cause analysis when an agent begins to underperform and provides the trend data needed to decide whether to adjust prompts, add knowledge, or refine actions. The dashboards turn raw traces into actionable signals rather than leaving you to review individual sessions manually.

A dashboard might show that resolution rate dropped from 78 percent to 61 percent over two weeks. Drilling into the underlying traces reveals that a new knowledge article introduced conflicting instructions. The Optimizer flags the article and suggests rephrasing the conflicting section. After the update, a follow-up test run confirms the resolution rate returns to the prior level.

Dashboards also surface cost signals. When an agent repeatedly calls an expensive external service, the metric appears as elevated cost per conversation. The trace identifies the exact action responsible, allowing you to replace the call with a cached internal lookup. These signals only appear once enough sessions accumulate, so early deployments require manual trace review until the data volume supports reliable trends.

When is this level of observability the wrong choice

If your agents perform only simple, deterministic tasks with no external data sources and no compliance requirements, the overhead of enabling and reviewing traces adds little value. In those cases lighter logging inside existing Flows or Apex usually suffices.

How do you enable tracing and testing in your org?

Start in Setup by turning on Agentforce Session Tracing, then add Agent Platform Tracing and Optimization features as needed. Assign the required permissions such as Data 360 User and Tableau Next Limited Consumer. Test changes in a sandbox rather than production when actions modify data.

In Agentforce Builder, use Preview in Simulate mode for quick iteration without side effects, then switch to Live Test for end-to-end validation. Move to the Testing Center to run batch scorers at scale. Enable citations and verification subagents where identity or source proof matters. After deployment, review Command Center dashboards regularly and adjust based on the Optimizer suggestions. Official Salesforce Help articles on observability, Session Tracing, and the Trust Layer remain the authoritative reference for the latest configuration details.

This framework supports steady improvement because each cycle of test, deploy, trace, and analyze produces evidence you can act on directly.

Frequently asked questions

How long do session traces remain available for review?

Traces are stored according to your Data 360 retention policies and can be queried like any other object in the platform. You control access through standard permission sets.

Can external monitoring tools consume Agentforce telemetry?

Yes. The platform exposes traces through OpenTelemetry, so Datadog, Splunk, and similar systems can ingest the data alongside your other application logs.

Do citations prevent every instance of hallucination?

Citations reduce the risk by tying responses to explicit sources, but they do not eliminate the possibility of incorrect synthesis. Review traces and scorer results to catch remaining issues.

What editions support the full set of tracing features?

Enterprise, Performance, Unlimited, and Developer Editions qualify when Salesforce Foundations or Agentforce 1 Edition is present. Certain tracing capabilities are not available in Government Cloud.

How do you measure whether the added observability justifies the setup effort?

Track the time saved during debugging and the reduction in post-deployment incidents. Organizations typically see the largest return when agents handle variable or regulated workflows that require compliance trails.

Book an intro call.

Found this useful? Pass it on.

Share it with someone on your team who's working through the same problem.

Funnelists Services

Agentforce & AI Agents

Autonomous agents that predict, automate, and execute revenue-driving actions across sales, service, and marketing.

Troy Amyett

Troy Amyett

Founder & Chief Solutions Architect

9x Salesforce certified. Agentforce Specialist and Agentblazer Legend, 2025–2026. Anthropic-certified in Claude Code and MCP.

Get Insights in your inbox

AI-powered perspectives on Salesforce and Agentforce, delivered weekly.

No spam. Unsubscribe anytime.

Your competitors aren't waiting. Why are you?

Let's talk about what AI agents could do for your business. 30 minutes. No pitch deck. Just answers.

Launch before they do