In this article, we answer the most common questions about Botmaker 3.0, the agentic version of the platform. They are grouped by topic so you can quickly find what you are looking for.
It is the entry point of your project and the coordinator of your entire team of AI Agents. It receives messages from the channels, talks with the user, and decides which agent should resolve each query. It owns the voice and tone, and it is the only one that talks to the user.
The orchestrator serves a function similar to the Masterbot, but expanded: in addition to routing, it converses, maintains the tone, can combine several agents in a single response, and switch between them as the conversation unfolds.
Instead of building rigid decision trees, you define goals in natural language and the agent interprets them. The main association shifts from "bot + channel" to "channel + orchestrator," and you reuse agents and languages without duplicating bots.
No. The agents resolve tasks and return the result to the orchestrator; the one who responds to the user is always the orchestrator.
No. The boxes are the states (steps) that a single agent follows to accomplish its goal. An agent is an agent; its internal steps are not separate agents.
Yes. It can consult more than one agent and synthesize a single response from what each one contributes.
Yes. The orchestrator can switch between agents as the conversation evolves; it does not follow a fixed linear path.
It reads the description (activation) of each agent: a short natural-language sentence that indicates when it should be used. There is no need to program the routing logic.
No. The relationship is 1:1: a channel is associated with a single orchestrator, so conversations don’t get mixed up. A single orchestrator, however, can have several channels connected.
Email linking is done through Mailbot; then you return to AI Agents to continue with the configuration.
This is determined in the orchestrator’s “Tone and style” section. You can request that it speak a specific language or that it adapt to the language the user writes in: if they write in English, it responds that way; if they write in Spanish, it responds in Spanish, and so on. It is not guided by the phone number’s prefix, but by how the user speaks to it.
Yes. A single orchestrator can handle different languages, without needing to build one bot per language.
Yes. Agents can be reused across several orchestrators without duplicating them, which avoids redoing processes.
There is no maximum number: an orchestrator can have as many agents as you need.
When you want the same team of agents to behave differently depending on the channel (for example, an orchestrator for Instagram with a casual tone and another for email that is more formal), or to separate large areas of your operation.
This function doesn’t exist at the moment; it is a planned improvement. In the meantime, configuration is done manually.
By entering the orchestrator and using the pencil icon that appears next to the title, at the top.
The orchestrator is deleted from the option at the end of its configuration ("delete orchestrator"). An agent is deleted from its profile, next to the workflow view.
Three: instruction (specifying a task in natural language), conditional (opening paths based on a condition), and loop (repeating steps over a set of items until a condition is met).
By adding required data in the corresponding workflow state. The agent will ask for it in order to continue.
Yes. If you prefer it to ask for the data one item at a time, conversationally, indicate this in the orchestrator’s tone and style.
You can test a piece of workflow logic in isolation. To test the complete behavior, that is done from the orchestrator.
Documents such as PDF, DOC, XLSX, or TXT, or else the URL of a web page.
The upload is processed asynchronously (indexing). The base becomes available to the agent only once its status changes to "ready."
By re-uploading the source material when it changes, so the agent always responds with current information.
Integrating means linking an external service’s account to the project (done once). Connecting means enabling that integration within an agent, choosing its actions and describing when to use them.
The agent can report that it cannot complete the action temporarily, because integrations depend on the external service.
An external server from the project’s catalog that exposes additional tools for the agents to use.
Not for the conversational part: the bot has memory and remembers the context automatically. Variables are reserved for integrations and code actions.
Not on its own: memory is per channel. To recognize a person across channels, you need an external system (CRM or spreadsheet) that uses a data point as an identifier.
No. The ticket is unique and moves along through the processes, regardless of which agent participates.
By adding a "route to a human" step in the agent’s workflow. It is not automatic: it is configured, and it can be assigned to a user or to a team.
At this stage, they participate as a validator or approver within the process; the conversation with the user is still managed by the AI Agent. This functionality is evolving.
They arrive as cards in the tickets section, alongside tickets from previous versions (legacy tickets).
From the orchestrator’s Test button. Testing the orchestrator is equivalent to testing the complete bot. You cannot test an agent separately; but you can test a piece of workflow logic in isolation.
In the tester (with "View details" you can see which agent it routed to) and in the Executions section. That reasoning is not shown in the conversation the end user sees.
The project’s authorized profiles (for example, those responsible for designing the agents).
By typing the @reset command, which clears the context and lets you start from scratch.
No, they are saved automatically. In addition, several people can edit at the same time, in real time.
Not at the moment: the generative model comes preconfigured.
By tokens, the unit by which the model’s processing is measured. You can review your consumption in the Products and consumption section, in the generative consumption view, with breakdowns by project, by model, and by function.
With concise designs: a brief tone and style and clear instructions make the model process less information.
Remember to visit our Help Center for further information.