Review Agentforce Coworker Prompts and Responses in Data 360

Query the Einstein Audit and Feedback data model objects (DMOs) in Data 360 to see what your users ask Agentforce Coworker, what Coworker sends to the LLM, and how the LLM responds.

REQUIRED EDITIONS AND PERMISSIONS
Available in:	Enterprise, Unlimited, and Agentforce 1 Editions
To query Agentforce Coworker audit trail data in Data 360:	See permissions table in Create and Run Queries in Query Editor.

Before You Begin

Turn on Einstein Audit and Feedback data collection. See Einstein Audit and Feedback Data.

Query the Agentforce Coworker Audit Trail in Query Editor

Each audit trail query joins three DMOs: GenAIGeneration to GenAIGatewayResponse on generationResponseId__c, and GenAIGatewayResponse to GenAIGatewayRequest on generationRequestId__c = gatewayRequestId__c. To filter for Agentforce Coworker, use req.feature__c = 'Agentforce Coworker'.

In Data 360, open Query Editor.
Enter one of the example queries in this topic.
Run the query.

Example Queries for Agentforce Coworker Audit Trails

List All Agentforce Coworker LLM Calls by User

Return the most recent 100 Agentforce Coworker LLM calls, with the session, user, model, full prompt, and raw response for each call.

1-- all Coworker LLM calls by user
2SELECT
3   req.timestamp__c                                                AS event_time,
4   req.sessionId__c                                                AS session_id,
5   req.UserId__c                                                   AS user_id,
6   gen.feature__c                                                  AS feature,
7   req.model__c                                                    AS model,
8   req.prompt__c                                                   AS sent_to_llm,
9   gen.responseText__c                                             AS llm_response
10FROM        GenAIGeneration__dlm gen
11JOIN        GenAIGatewayResponse__dlm res
12   ON      gen.generationResponseId__c  = res.generationResponseId__c
13JOIN        GenAIGatewayRequest__dlm req
14   ON      res.generationRequestId__c   = req.gatewayRequestId__c
15WHERE req.feature__c = 'Agentforce Coworker'
16ORDER BY req.timestamp__c DESC
17LIMIT 100

Extract the Agentforce Coworker Response Text

The llm_response column contains raw JSON, with tool invocations mixed in with the response text. This query extracts only the text content.

1-- Extracting just the LLM response TEXT
2SELECT
3   req.timestamp__c                                             AS event_time,
4   req.UserId__c                                                AS user_id,
5   SUBSTRING(
6       gen.responseText__c,
7       POSITION('"content":"' IN gen.responseText__c) + LENGTH('"content":"'),
8       POSITION('"' IN SUBSTRING(
9           gen.responseText__c,
10           POSITION('"content":"' IN gen.responseText__c) + LENGTH('"content":"')
11         )) - 1
12   )                                                            AS llm_response_text
13FROM        GenAIGeneration__dlm gen
14JOIN        GenAIGatewayResponse__dlm res
15   ON      gen.generationResponseId__c  = res.generationResponseId__c
16JOIN        GenAIGatewayRequest__dlm req
17   ON      res.generationRequestId__c   = req.gatewayRequestId__c
18WHERE req.feature__c = 'Agentforce Coworker'
19 AND POSITION('"content":"' IN gen.responseText__c) > 0
20 AND SUBSTRING(
21   gen.responseText__c,
22   POSITION('"content":"' IN gen.responseText__c) + LENGTH('"content":"'),
23   1
24   ) != '"'
25ORDER BY req.timestamp__c DESC
26LIMIT 100

Extract the Dynamic Context from Agentforce Coworker Prompts

The system instructions are the same in every Agentforce Coworker prompt. The ## Dynamic Context section makes each prompt unique, with details such as the user’s identity, recently viewed records, and session state. This query returns each prompt from the ## Dynamic Context marker onward.

1-- Query to extract from ## Dynamic Context onwards
2SELECT
3   MIN(req.timestamp__c)                                        AS event_time,
4   req.UserId__c                                                AS user_id,
5   SUBSTRING(
6       req.prompt__c,
7       POSITION('## Dynamic Context' IN req.prompt__c)
8   )                                                            AS dynamic_context_onwards
9FROM        GenAIGeneration__dlm gen
10JOIN        GenAIGatewayResponse__dlm res
11   ON      gen.generationResponseId__c  = res.generationResponseId__c
12JOIN        GenAIGatewayRequest__dlm req
13   ON      res.generationRequestId__c   = req.gatewayRequestId__c
14WHERE req.feature__c = 'Agentforce Coworker'
15 AND POSITION('## Dynamic Context' IN req.prompt__c) > 0
16GROUP BY req.prompt__c, req.UserId__c
17ORDER BY event_time DESC
18LIMIT 100

Extract the Agent Catalog from Agentforce Coworker Prompts

In the ### Agent Catalog section, Agentforce Coworker uses the dynamic context to decide which agent or tool to call. This query returns each prompt from the ### Agent Catalog marker onward, so you can review the routing decision.

1-- Query to extract from ### Agent Catalog onwards
2SELECT
3   MIN(req.timestamp__c)                                        AS event_time,
4   req.UserId__c                                                AS user_id,
5   SUBSTRING(
6       req.prompt__c,
7       POSITION('### Agent Catalog' IN req.prompt__c)
8   )                                                            AS agent_catalog_onwards
9FROM        GenAIGeneration__dlm gen
10JOIN        GenAIGatewayResponse__dlm res
11   ON      gen.generationResponseId__c  = res.generationResponseId__c
12JOIN        GenAIGatewayRequest__dlm req
13   ON      res.generationRequestId__c   = req.gatewayRequestId__c
14WHERE req.feature__c = 'Agentforce Coworker'
15 AND POSITION('### Agent Catalog' IN req.prompt__c) > 0
16GROUP BY req.prompt__c, req.UserId__c
17ORDER BY event_time DESC
18LIMIT 100

Extract What Users Asked Agentforce Coworker

The role: user section of a prompt contains the generalized, rewritten version of the user’s question. This query extracts this text so you can see which tasks and topics drive Agentforce Coworker usage.

1-- Query to extract only the generalized query of what the user asked
2SELECT
3    MIN(req.timestamp__c) AS event_time,
4    req.UserId__c AS user_id,
5    SUBSTRING(
6        req.prompt__c,
7        POSITION('role: user' IN SUBSTRING(
8            req.prompt__c,
9            POSITION('### Agent Catalog' IN req.prompt__c)
10        )) + POSITION('### Agent Catalog' IN req.prompt__c) + LENGTH('role: user'),
11        GREATEST(
12            POSITION('role: assistant' IN SUBSTRING(
13                req.prompt__c,
14                POSITION('role: user' IN SUBSTRING(
15                    req.prompt__c,
16                    POSITION('### Agent Catalog' IN req.prompt__c)
17                )) + POSITION('### Agent Catalog' IN req.prompt__c) + LENGTH('role: user')
18            )) - 1,
19        0)
20    ) AS user_question
21FROM        GenAIGeneration__dlm gen
22JOIN        GenAIGatewayResponse__dlm res
23    ON      gen.generationResponseId__c  = res.generationResponseId__c
24JOIN        GenAIGatewayRequest__dlm req
25    ON      res.generationRequestId__c   = req.gatewayRequestId__c
26WHERE req.feature__c = 'Agentforce Coworker'
27    AND POSITION('### Agent Catalog' IN req.prompt__c) > 0
28GROUP BY req.prompt__c, req.UserId__c
29HAVING
30    -- Ensure 'role: user' exists after '### Agent Catalog'
31    POSITION('role: user' IN SUBSTRING(
32        req.prompt__c,
33        POSITION('### Agent Catalog' IN req.prompt__c)
34    )) > 0
35    -- Ensure 'role: assistant' exists after 'role: user'
36    AND POSITION('role: assistant' IN SUBSTRING(
37        req.prompt__c,
38        POSITION('role: user' IN SUBSTRING(
39            req.prompt__c,
40            POSITION('### Agent Catalog' IN req.prompt__c)
41        )) + POSITION('### Agent Catalog' IN req.prompt__c) + LENGTH('role: user')
42    )) > 0
43ORDER BY event_time DESC
44LIMIT 100

The prompt-extraction queries return reliable results for primary conversational calls, which contain the ### Agent Catalog marker. Delegated agent calls, internal helper calls, and utility calls don’t contain the same markers, so rows for those calls can return partial fragments, --, or empty results. The POSITION('### Agent Catalog' IN req.prompt__c) > 0 filter removes most of those rows, but some can remain.

Note

To learn more about the prompt sections and LLM call types, see Agentforce Coworker Data in Data 360.