Skip to main content
Waaru
WhatsApp AI agents

How WhatsApp AI agents use tools and take actions

Tanay PatelAI and full-stack engineerUpdated 22 October 2026
The short answer

A tool gives a WhatsApp AI agent a defined way to read information or perform an action in another system. The application checks that the request is allowed, runs the action and gives the result to the agent. A convincing reply does not prove an action succeeded; the business system must confirm the result.

“Where is my order?” sounds like a simple question. Answering it requires the right order, access to its current status and a reply that says only what that status supports. The example below explains the process. It uses no real customer data and does not describe a specific integration.

How is a tool different from a knowledge base?

A knowledge base can explain your dispatch policy. A tool can check an order’s current status in your order system, provided the customer is allowed to see it. Those are different sources with different jobs.

A policy saying that orders usually leave the warehouse promptly is not evidence that this customer’s parcel has shipped. If the live record is unavailable, say that the status could not be checked.

Keep general information in a reviewed knowledge base. Check the integration’s documented actions for actions involving an individual customer’s records.

What happens during an order lookup?

From an order question to a checked answer
  1. Customer asks about an order

    Ask which order if the reference is missing or unclear. DEMO-ORDER is a made-up reference for this example.

  2. Check the order and permission

    Check within the correct business’s records. Confirm the customer is allowed to see the order; its reference alone is not enough.

  3. Look up the current status

    Read the connected order system. Check the information it returns before writing the reply.

  4. Reply with the confirmed result

    If the status is “processing”, say that. Do not promise “it arrives tomorrow” without delivery information to support it.

If a check fails, pause. Explain what could not be checked. Ask for the missing detail or offer help from a person instead of guessing the order status.

How is changing a record different from reading it?

Developers call fetching information a read and changing it a write. Looking up a booking and moving it to another date need different permissions and checks.

For a change, confirm the intended record, the requested change and any required approval. Then check the business system’s response. Telling a customer “your booking has moved” before confirmation creates a promise the system may not have kept.

Some tasks belong with a person, especially when someone needs to approve a price or an exception. The sales workflow leaves a custom quote with the teammate who decides what work the quote includes.

What if the tool fails or the result is uncertain?

Decide what the customer hears for each failure. Try these situations in your chosen system. They are suggested checks, not Waaru test results.

  • No matching record: ask for the necessary clarification without exposing other customers’ records.
  • Access denied: do not switch to a more powerful account to get around the restriction. Follow the business’s identity-check or support process.
  • Outdated information: explain when the status was last checked. Do not present an old update as current.
  • No response after a change: the action may have completed even though the reply was lost. Check what happened in the source system before trying again.
  • Unsupported request: explain the limit and offer a person. Do not substitute a different action simply because that tool is available.

Where does MCP fit?

The Model Context Protocol, or MCP, gives AI applications a shared way to discover and call tools. The MCP tools specification describes tool names, the information each tool needs, its results and how it reports errors.

MCP is not a guarantee of a particular action. Each server offers its own set of tools. The application must still check permissions and the details supplied, and ask a person to approve an action when required. A tool’s description does not replace these checks.

The WhatsApp MCP guide explains how this relates to WhatsApp. Before planning a workflow, check that the service is running and supports the action you need.

Is the Waaru SDK the same thing as an agent tool?

No. The Waaru SDK for Node.js helps developers call the public API from their server. Its methods send messages, read conversations and delivery status, stream permitted media, and manage or verify callbacks. It does not provide order lookups or MCP tools.

Use the Node SDK guide and the official @waaru/sdk package for supported methods and examples. Published SDK beta.4 covers 32 operations. The REST API also has newer operations and fields; the SDK guide explains which version supports each one. Check each method’s required scope and keep its key on your server.

What should you verify before connecting a tool?

Choose one action and try it with made-up data in a test environment. Record what you expected, what happened, which system supplied the result and what the customer would hear if it failed. Never include secrets or customer records in a public test report.

Use the AI-agent test guide for the evaluation sheet and the handoff guide for requests a person must own.

Before adding more tools, check whether a fixed flow or an agent is the right fit. Check whether the customer’s task actually completed.

FAQ

Frequently asked questions.

Does an AI agent need a tool for every answer?

No. A reviewed policy can answer a general question. A live lookup or business action needs the appropriate connected operation and permission checks.

Does MCP provide every WhatsApp integration automatically?

No. MCP describes how tools are exposed and called. Each server and connector implements its own operations, limits and authentication.

Should an agent repeat a timed-out action?

Not blindly. A timeout can leave the outcome uncertain, especially after a write. Check the source system or follow its instructions for checking a failed or uncertain action before attempting the change again.

Research notes

Evidence checked for this article.

  • MCP defines discoverable tools, schemas, results, errors and security considerations; actual operations depend on the server implementation.

    Model Context Protocol, 2025-06-18 specification · Accessed 15 September 2026

About the author

Tanay Patel

AI and full-stack engineer

Tanay builds production AI systems, agent workflows, and the full-stack software around them. His work focuses on reliable state, recovery, and human handoff.

Read Tanay's engineering notes

Put this into practice on your own number.

The platform is free. AI features are paid separately.