Tracing sessions of Boomi Agent Garden agents

Agent session logs provide comprehensive visibility into agent behavior, performance, and troubleshooting details. They help you understand how an agent reasons, diagnose issues quickly, and optimize agent effectiveness. Refer to Testing and troubleshooting your agent to learn more.

Session logs are only available for Boomi Agent Garden agents. They allow you to:

View an agent’s reasoning while running in production mode.
Track and review all sessions executed by the agent.

Key benefits of session logs

Transparency into agent decision-making: Using the chain of thought reasoning, session logs provide real-time visibility into the agent’s thought process and allow you to understand its step-by-step logic. This feature allows you to understand why an agent produced a specific recommendation or result.
Streamlined debugging and faster resolution: You can access comprehensive logs that include tool inputs and outputs, LLM reasoning, and full invocation traces. Detailed error logs and invocation-level debugging enable you to quickly identify root causes at the tool level, significantly speeding up the troubleshooting process. Session logs provide clear, understandable explanations of what went wrong and which specific step in the process failed.
Comprehensive performance monitoring: The logs provide essential performance data, including Time to First Token (TTFT), average latency, and success/failure rates per session or invocation. You can track token consumption (input and output) and the number of LLM invocations, which is critical for managing costs in high-volume workflows. This comprehensive monitoring is done at the agent, session, and invocation levels.
Governance and compliance: You can monitor when and why specific a guardrail is triggered, identifying policy violations or content types that were blocked or warned.

Navigating to Session logs

You can access session logs from either Agent Garden or Agent Control Tower.

Through Agent Garden

While conversing with an agent

When chatting with an agent in production:

Under the agent’s response, click the View log icon.
You are redirected to the current session log in Agent Control Tower.

Session logs from chat

From the agents page

Navigate to the required agent.
Click the Actions menu.
Select View log to open the agent’s session log page.

Through Agent Control Tower

From the Agents page

Select the required agent.
Click View Details.
Navigate to the Logs tab.

From the Boomi Agent Garden provider account

Navigate to Manage Providers and select the Boomi Agent Garden provider.
Select the provider account containing the agent.
Open the Agents tab.
Select the required agent and click View Details.
Navigate to the Logs tab.

Understanding session logs

The session logs help you understand the agent's performance on three different levels:

The overview performance of the agent(Account level): You can see agent performance on account level, i.e., aggregate of agent metrics by all users of the account.
The in-depth session metrics of the agent(User-specific): Session metrics are user-specific, showcasing their interactions with the agent.
An in-depth view per invocation the agent has made(User-specific): Invocation metrics are user-specific, showcasing their interactions with the agent.

Overview level

You can filter session logs using the time filter. Based on the selected time range, the Overview page displays aggregated agent metrics and allows navigation to individual sessions.

Agent metric	Description
Total Sessions	Total number of agent interaction sessions initiated with the agent.
Error Rate	Percentage of invocations where the agent encountered execution errors or failed to complete successfully.
Guardrails Triggered	Count of times safety, policy or validation guardrails were activated during agent execution.
Tool Calls	Tracks frequency of automated tool executions triggered by the agent to fulfill user requests beyond simple text generation.
Average Latency	Average time taken by the agent to process and respond per interaction.
Total Tokens	Total number of tokens consumed by the agent for both input and output during execution.

Under the Sessions section, you can see details in regard to:

Metric	Description
Session Name/ID	Brief name describing what the invocation was about, along with the Session ID present under the name.
Last Invocation	Date and time of when the invocation occurred.
Invocation	The number of times the invocation was carried out.
Guardrails Triggered	If any guardrails are triggered. Options are Yes or No.
Tool Calls	Tracks the frequency of automated tool executions triggered by the agent to fulfill user requests beyond simple text generation.
Error Rate	Indicates if an error occurred in terms of percentage. This percentage us calculated, based on the total number of invocations and the number of successful invocations. For example, supposing for every 100 invocations in one session, 87 invocations fail, then error rate is 87% and success rate is 13%.

Session level

Clicking a Session Name/ID opens the Session page.

The Invocation Success Rate graph provides a visual breakdown of successful and failed invocations.

The following Session Metrics are recorded:

Metric	Description
Invocation	Invocation is the number of user prompts/conversations started with the agent within a single session.
Average Latency	The average time taken by an agent in a session to respond to user prompt.
Guardrail Hit	This is a Yes/No indicator that tells if a guardrail was hit for a particular invocation. Guardrail hit is a terminal event, it tells you that the invocation was a failure.
Total Tokens	Total token consumed by the agent, this includes input tokens and output tokens.
Tool Calls	Tracks the frequency of automated tool executions triggered by the agent to fulfill user requests beyond simple text generation.
Session Duration	The total elapsed time from your initial prompt until the final response or session timeout, reflecting the complete lifecycle of the interaction.

The Invocations table provides the following details:

Metric	Description
Time	Date, time, and time zone of when the invocation occurred.
User Input	The prompt that was provided by you.
Agent Response	The agent's response to the prompt.
Status	If the invocation was a success or a failure.
Guardrails	If any guardrails were triggered.

Invocation level

Clicking an invocation opens a detailed view with the following information:

A summary section covers:

Metric	Description
Agent	Name of the agent.
Trace ID	ID assigned to that particular invocation.
Guardrails Hit	This is a Yes/No indicator that tells if a guardrail was hit for a particular invocation. Guardrail hit is a terminal event, it tells you that the invocation was a failure.
Error	The number of errors faced in that particular invocation.
Total Steps	Number of steps the chain of thoughts took to respond back to the prompt.
Tool Calls	Tracks the frequency of automated tool executions triggered by the agent to fulfill user requests beyond simple text generation.
LLM Performance metrics	Performance metrics of the Large language model, which include: Total Invocation Time:The amount of time an invocation took to occur. LLM Invocations: The number of times the LLM responded with an answer. Total Tokens: Total token consumed by the agent, this includes input tokens and output tokens. Input Tokens: Total number of input tokens consumed. Output Tokens: Total number of output tokens consumed. Time to First Token: Measures the delay between the prompt and the very first token response. Average Latency: The average time taken by an agent in a session to respond to user prompt.

You also get to view the User Prompt along with the Agent Response for that particular invocation.

For an in-depth understanding of the processes behind the agent's response, the Chain of Thoughts runs you through the LLM thinking process, the tools called and its parameters.