← All stories
Product update

Introducing Agent Builder: build, test, and deploy your support agent

Learn what each part of Agent Builder does, which controls to use, and how to take a support workflow from configuration to a live customer channel.

By the Chatropic team8 min read
Agent Builder: the sub-agent workspace, shown with illustrative data.

Agent Builder is the workspace for configuring your support agent: its instructions, knowledge sources, connected tools, procedures, and customer channels.

Open your agent in the builder. Use this guide to configure each capability, check it in the Playground, and release the saved changes when they are ready.

Choose the capability your task needs

  • Agents: define who handles a request, their instructions, and the capabilities they can access.
  • Knowledge base: provide policies, documentation, and other approved sources for factual answers.
  • Data connectors: retrieve selected fields from a read-only API, such as a customer’s current cart.
  • MCP tools: expose operations from a connected business system and grant them to the appropriate specialist.
  • Skills: describe when a procedure applies and the steps, conditions, and handoffs it should follow.
  • Playground and Deploy: test the saved configuration, inspect activity, and configure where customers can use the agent.

For example, a delivery policy belongs in knowledge. A current order status requires a live lookup. A procedure that collects an order number, performs that lookup, and handles a missing order belongs in a skill.

Configure agents and instructions

Open the specialist you want to configure in the agent workspace. Its routing description explains when it should receive a request; its instructions explain how to handle that request.

  1. Give your custom specialist a clear name and routing description, such as handling order status and delivery questions.
  2. In Instructions, state the job, required evidence, fallback behavior, and escalation rules.
  3. Choose a Response tone: preserve the agent voice, use the brand voice, or apply per-message guidance.
  4. Add Suggested prompts under Conversation starters, with one customer-facing prompt per line.
Example instructions

Help customers track their orders. Ask for an order number before looking up an order. Use the lookup result for status and delivery information. If the order cannot be found, ask the customer to check the number. Hand off unresolved requests to support. Never invent a delivery date.

Some system specialists have platform-managed instructions. When those fields are locked, configure the available procedures, tool grants, prompts, and knowledge access instead.

Give a specialist a clear role, instructions, and boundaries.

Illustrative demo
Agent design: configure the specialist’s responsibility and instructions. Illustrated with example data.

Attach knowledge and connect live data

Knowledge base: answers from approved content

Add your policies and documentation in Knowledge base, then check their processing status. In the specialist’s Agent knowledge section, search Available knowledge and choose Attach beside each relevant source. Adding a source to the shared store and attaching it to a specialist are separate steps.

Use knowledge for information such as delivery areas, return conditions, and troubleshooting instructions. Keep those sources current, then test a question whose answer you can verify against the source.

Knowledge base: manage sources, then attach the relevant knowledge to a specialist. Illustrated with example data.

Data connectors: retrieve current information from an API

Use Data connectors when the answer depends on a live, read-only lookup. The connector editor supports GET requests; it is not a way to update an order or issue a refund.

  1. Enter a Connector name, Description, and API endpoint. Add any required Query inputs.
  2. Configure authentication and select the Allowed response fields, one field path per line. This controls which parts of the response the connector returns.
  3. Enable AI Agent triggered and fill in When should the AI Agent use this connector? Explain when to call it and how to use the result. Enable Human triggered if teammates should run it manually.
  4. Try a Mock response to check the output, then Test live connection to verify the endpoint and authentication. Save and test before enabling a channel.

For customer-specific data, configure the appropriate customer authentication and audience. Connector releases are managed from their own editor.

Connect MCP tools and build skills

MCP tools: connect a system and assign its operations

MCP is the connection between your agent and tools exposed by another system. The available operations depend on the server you connect; an order lookup is only available if that server exposes one.

  1. Open the MCP connection setup in Tools. Enter a Connection name and the server’s Streamable HTTP URL.
  2. Choose Connect and discover. Complete sign-in if requested, or configure the credential reference your server requires.
  3. Review the discovered tools. In the specialist’s MCP tools section, grant access to the operations it needs.
  4. Test a request that should invoke the tool and inspect the result. Connecting a server alone does not describe when a specialist should use each operation.
MCP tools: connect a server, discover its operations, and grant access to a specialist. Illustrated with example data.

Skills: build the procedure around those tools

In Associated procedures, choose Add skill. Set the procedure name and When to use conditions, then use Add step in the procedure editor. The editor provides these step types:

  • Instruction: describe what the agent should ask, check, or explain.
  • Use MCP tool: select the tool operation needed at that point in the procedure.
  • Condition: define a check and the alternative behavior when it is not met.
  • Section header: give a group of steps a named destination.
  • Go to section: continue at the specified section.
  • Handoff: route the request to the configured handoff target.
  • Switch skill: continue with another skill.
  • End procedure: stop the current procedure.
Example: an order-status skill

Set When to use to “The customer asks where an order is.” Add an Instruction to collect the order number, a Use MCP tool step for your connected lookup, and a Condition to check whether an order was returned. Explain the returned status when found; otherwise ask for a corrected number or use Handoff. Finish with End procedure.

The lookup name and returned fields must match your connected tool. Enable the procedure when it is ready for routing. Use the YAML view to inspect the generated procedure definition.

Turn order tracking into a reusable procedure with a connected tool.

Illustrative demo
Skills: add instructions, tool calls, conditions, and handoffs to a procedure. Illustrated with example data.

Test responses and tool activity

The straightforward question is only one part of a support conversation. A customer may leave out an order number, ask about a different order, or request an exception to your policy. Open the Playground (Test agent) and try the questions your team hears every day. Review the response and the activity trace to see which specialist, knowledge, and tools were used.

  • Can it answer a normal delivery question using your policy?
  • Does it ask for missing details before looking up an order?
  • What happens when the lookup fails or the customer asks for a refund?
  • Does it bring in a teammate when the request is outside its instructions?

Save your changes, then repeat the conversation in the Playground. Check the routing and tool activity: did the expected specialist respond, did the lookup receive the right input, and did the answer match its result? If no tool ran, review the specialist’s tool grants and the skill’s When to use conditions. Use test data while checking connected actions.

Try a customer question and follow the agent’s tool activity.

Illustrative demo
Test agent: try a customer question and inspect the tool activity. Illustrated with example data.

Enable and connect a channel

Save, test, then enable a channel

Save your agent settings and test drafts in the builder or playground Sandbox. Enabled channels use the saved configuration with live keys; agent publishing and sandbox API keys are no longer required. Data connectors have their own enablement controls.

Configure your customer channel

Open Deploy and choose All channels to see where customers can reach your agent. The page separates available channels from those coming soon, and shows which connections still need setup.

For a store’s website, choose Chat widget. Manage widget opens the editor where you can customize its appearance, choose allowed websites, and copy the installation code. WhatsApp and Slack have their own Connect actions; API access has a Manage keys option.

Set up your Chat Widget and explore upcoming channels such as Instagram.

Illustrative demo
Deploy → All channels: manage the Chat Widget, connect WhatsApp or Slack, and access API keys. Illustrated with example data.

Choose a channel that is available today and verify the setup before directing customers to it. Instagram and the other channels under Coming soon are not yet available to configure.

Review conversations and content gaps

Open Conversations to review real exchanges after launch. Check where a reply missed the question, a lookup failed, or a handoff was needed. Update the relevant instructions, tool access, or skill, and repeat that scenario in the Playground.

Unanswered questions can reveal what your knowledge base is missing. Open Reports → Content gaps to review those recommendations alongside the conversations that prompted them.

Select a recommendation to see its detection confidence, an article draft, and related conversation evidence. For an order-support agent, a gap might reveal that customers need clearer guidance on delivery exceptions. The example below shows the same review workflow for missing billing guidance.

Content gaps: review the recommendation, article draft, and related customer conversations together. Illustrated with example data.

Improve the source behind the answer.

Read the linked questions before deciding what to add. Choose Edit in knowledge editor to complete the draft and review the policy details. The draft is not published just because it appears here.

Use Mark as done when you have addressed the gap, or Dismiss if it does not need action. Open, Done, and Dismissed views help your team track the review work. Then test the original question again to check whether the updated knowledge helps the agent respond.

Your first working configuration

Before launch, check that your specialist has instructions, the right knowledge attachments, and any required connector or MCP tool access. Confirm the skill handles missing information and failed lookups. Save, test, set the required changes live, and verify your customer channel.

When an answer fails, change the part responsible: source content for a missing fact, a connector or tool for a failed lookup, and instructions or skill steps for incorrect behavior. Then test the same question again.