The difference between an agent that answers well and one that resolves lives in this layer: your own knowledge, tools that hit your systems, and a human handoff that genuinely works.
Last updated: September 2026
Meta Business Agent Platform is the API layer used to configure and operate the WhatsApp agent: you load knowledge into it, set brand instructions, declare your own APIs as tools, subscribe webhooks, and govern who controls each conversation, the agent or your application.
The platform covers WhatsApp only today, on the Cloud API. Switching the agent on from the WhatsApp Business app gives access to none of these surfaces.
Source: official Meta Business Agent Platform documentation, 2026.
Before the first API call, the boring part has to be settled, and that is where most integrations stall.
This is the platform's quietest failure and it costs hours of debugging.
Knowledge is not one giant prompt: it is four distinct surfaces, maintained separately and auditable.
This is where the agent's behavior is defined. It is the equivalent of system instructions, and it is what separates a brand agent from a generic bot.
A connector is the definition of an external API. Its tools are the concrete operations the agent may call during a conversation.
This is the most underestimated block. Without it, human escalation is a message asking the customer to call you.
The platform lets you validate the agent before it ever talks to a real customer. Skipping this is the fastest way to burn trust in the channel.
The sensible way to go live: the agent enabled, but talking only to whoever you decide.
If the integration serves several clients, the credential model changes and it is worth setting up properly from the first project.
The platform's real sequence, grouped by phase.
Paths per Meta's official documentation as of September 2026. The platform moves fast: always verify against the current reference.
The implementation service, the setup requirements and token-based cost.
Yes, if you are integrating for your own business: a Meta app with the right permissions and a system user is enough. The technology partner pattern becomes necessary when you operate third-party accounts and need credentials separated per client.
It depends on the tools you declare. A connector can expose read operations, such as looking up an order, and write operations, such as creating a booking. The surface is whatever you decide, and that is exactly where the limits belong.
From the webhooks. Messages the agent handles arrive on the standby field, and every control change is notified in its own event. Sending a message from your application takes control implicitly.
Not simultaneously. One number cannot run several messaging products at once. What you can do is split by number, or use conversation control so the agent and your application alternate on the same thread.
The agent loses the data it was about to fetch, so the failure strategy has to be designed: escalate to a person, answer with a safe fallback, or ask to retry. That is part of the integration design, not something the platform solves on its own.
Yes. There is an evaluation endpoint to measure performance and an insights endpoint that shows which knowledge is being used. That combination is what lets you fix bad answers instead of guessing.
Tell us which APIs you have and what the agent should resolve, and we come back with the connector design, conversation control model and test plan.
Integrate your API with the agent?
We will talk to your engineers