Build a support context agent

Support teams answer the same questions with different context every time: who this customer is, what they bought, what they asked before, what was promised. An agent that gathers that context before a human replies makes every answer faster and more consistent. Here is the shape of one.

What the agent does

For each new conversation, it prepares a side panel for the support person:

• a summary of the customer’s history and earlier tickets,
• relevant account and billing details,
• what was discussed by email, including attachments,
• a suggested reply the human can edit and send.

The services

Human in the loop by design

The agent drafts; a person sends. Enable read endpoints for context, and keep sending in your own support tool’s hands. If you later let the agent post replies directly, add that one write operation in staging first and limit it to the requests that need it.

The workflow

Configure a support_context capability that takes a customer email address and:

1. retrieves earlier tickets,
2. retrieves the customer record and subscription status,
3. returns one structured result.

Give the model iGPT as a separate tool for questions about past email threads, so it asks only when the history matters.

The request

Conceptual example.

const panel = await agent200
  .user(agent.id)
  .run({
    model: "openai/gpt-5.6-sol",
    instructions: `
      Prepare context for a support person.
      Summarize history in three bullets, then draft a reply in our tone.
      Cite the ticket or message behind every fact. Never promise refunds.
    `,
    input: ticket.latestMessage,
    tools: ["support_context", "email_context"],
  });

Make it consistent

• Put your tone and your policies in the instructions, the same for every request.
• Use a lighter model to classify and route, and a stronger one to draft.
• Track how much of each draft the team edits. Heavy edits point to instructions worth improving.

See more examples on the Use Cases page, or the services in the integrations catalog.