Email Conversations

Every email that reaches the agent and every response it sends is automatically recorded as a conversation. If the maintenance manager sent an email asking about the status of the north plant compressors, you can see exactly what the agent replied, which tools it used to obtain that information, and whether the agent processed any attachments. Nothing is lost.

What is it for?

When the agent handles hundreds of emails per month, being able to supervise the quality of those responses matters. The conversation history is used for:

• Supervision: verifying that the agent is responding correctly and with up-to-date information
• Diagnosis: if someone reports that "the agent gave an incorrect response," you can see exactly what happened and why
• Audit: demonstrating that a request was received and answered, with the exact timestamp
• Continuous improvement: identifying conversations where the agent struggled, to adjust its prompt or tools

How does it work?

Each email thread is stored as a conversation. When the same user sends a follow-up reply to the agent's email, the system recognizes it as part of the same thread and maintains the full conversation context — the agent remembers what was discussed earlier and can continue coherently.

The system also records tool calls — every time the agent consulted a tool to retrieve data or execute an action. Viewing them allows you to understand the reasoning behind each response.

How to use it?

View the conversation list

Go to Email > Conversations in the sidebar. The list shows all conversations from all your email agents:

View a conversation in detail

Click any conversation to see the full thread:

• User messages: the emails that reached the agent, with the subject, sender, and full body
• Agent responses: the emails sent by the agent, exactly as the user received them
• Processed attachments: the PDFs, images, or documents the agent analyzed using AI vision, with the extracted text or data
• Tool calls: the tools the agent used to build the response — what it queried, with which parameters, and what it received

Read tool calls to understand the agent's reasoning

Tool calls are the agent's "invisible work" — the queries and actions it executed before writing its response. Viewing them lets you know exactly why the agent replied what it did.

Example of a visible tool call in a conversation:

• The supplier sends an email: "Please confirm receipt of invoice INV-2891 for $45,200."
• The agent executes: Tool "query_invoices" with parameter number = "INV-2891"
• The tool returns: invoice record, amount, receipt date, payment status
• The agent responds: "We confirm receipt of your invoice INV-2891 for $45,200, registered on March 15. Payment is scheduled for March 30 per our 15-day payment terms."

If the response had been incorrect, viewing the tool call would show you whether the problem was the tool returning wrong data, or the agent misinterpreting correct data.

Search for a specific conversation

Use the filters in the list to find a conversation:

• Filter by sender (the user's email address)
• Filter by date range
• Filter by agent (if you have multiple email accounts with different agents)

Key benefits

• Permanent history of all agent emails — no conversation is lost
• Full visibility into the agent's reasoning through tool calls
• Audit trail with exact timestamps for every message sent and received
• Record of processed attachments for verifying data extraction
• Quick identification of problematic conversations to improve the agent
• Previous message history is maintained for coherent responses in long threads

Common use cases

Scenario 1: Supplier request audit A supplier claims they sent their technical proposal by email but never received a response. The procurement manager opens the conversation history and searches for the supplier's email address. They find the March 12 conversation: the supplier sent the PDF proposal, the agent processed it, and responded within 4 minutes confirming receipt and providing the assigned reference number. The manager shares a screenshot of the conversation with the supplier as evidence.

Scenario 2: Incorrect response diagnosis The maintenance manager reports that the agent gave information about the wrong equipment when asked about compressor C-03. The supervisor opens that conversation in the history, views the query tool call, and finds that the agent searched for "C03" without a hyphen in the collection and retrieved the record for "Heater C03-A" instead of the compressor. They correct the prompt instruction: "Always search for equipment with the hyphen in the code (C-03, not C03)."

Scenario 3: Weekly quality supervision The coordinator reviews the last 20 conversations of the week every Friday. They read the longest threads (those with more than 3 messages) to verify the agent is maintaining context correctly. They notice that in one thread the agent forgot to mention the task number created in the second message. They update the prompt: "When creating a task, include the task number in every subsequent response within the same thread."