Agent Workflows Route Conversations Through Specialized Subagents

ElevenLabs presents Workflows as a visual editor for building conversational agents around routed paths, specialized subagents, and explicit handoffs rather than concentrating all business logic inside a single agent. The product is framed as a step toward “agentic control”: builders define when an agent should interpret a request, when routing logic should apply, when a predefined action should run, and when a conversation should move to a human operator.

The clearest example is a routing agent that classifies an incoming caller’s intent and sends the conversation to a specialized branch. The displayed routing node is instructed to “Determine if the enquiry is technical or pricing-related.” One condition, “The caller has a technical question,” leads to a “Technical Help” node that asks and guides the user through troubleshooting. Another condition, “The caller has a pricing question,” leads to a “Pricing Enquiry” node that answers the caller’s questions using a provided knowledge base.

That example defines the basic model. A workflow is not presented as a linear phone tree, but it is also not a single open-ended conversational agent. It is an orchestrated structure in which AI interpretation, routing logic, and predefined actions can sit on the same canvas.

ElevenLabs’ product language groups those elements as “artificial intelligence,” “complex logic,” and “deterministic action.” In the material shown, those categories appear through examples rather than a formal architecture diagram: routing a caller by question type, dispatching a shipment, verifying a customer, and transferring a prospect to sales. The point is not that every step is left to a model. The editor is positioned as a place to combine conversational AI with explicit business paths and actions that builders define in advance.

Workflows are built around specialized subagents. Instead of giving one agent every instruction, policy, knowledge source, and tool, the workflow can route a user to subagents with narrower responsibilities. ElevenLabs says each subagent can draw from its own tools, knowledge base, or language model, with the goal of optimizing “cost, latency, and accuracy at every stage.”

The interface shown for a new subagent makes that separation concrete. The configuration panel includes tabs or sections for “General,” “Knowledge Base,” “Tools,” and “Agent Testing.” It also shows options to include a global knowledge base, add an additional knowledge base document, define a “Conversation Goal,” set a voice, and select an LLM. In the displayed configuration, the voice is “Vincent - Deep and Relaxing,” and the selected model is “Claude Sonnet 4.”

The “Conversation Goal” field is described in the interface as a way to “extend the global prompt with information specific to this conversation node.” That points to a layered design: the workflow can carry global instructions, while individual nodes add local instructions for the immediate task. A pricing node can be grounded in pricing material. A technical-support node can use troubleshooting knowledge. A sales-transfer node can focus on routing the potential client to the relevant team.

The practical claim is that narrower subagents let builders scope context and capability around the current step in the conversation. A workflow does not require every exchange to use the same prompt, the same knowledge base, the same tools, or the same model. Each stage can be configured for its particular job.

The workflow elements shown are not limited to answering user questions. ElevenLabs positions the editor as a way to connect conversation design with operational steps: triggering real-world events, verifying and authorizing callers, and transferring conversations to a human operator.

The examples are concrete. A “Dispatch Shipment” node is described as: “Trigger sorting and delivery of product to the customer's location.” A “Verify Customer” node searches for the user’s contact information within a knowledge base. A “Transfer to Sales” node redirects a potential client to the sales team in their region.

Those nodes expand the product’s scope beyond support-chat response generation. The same flow can include technical help, pricing answers, customer verification, shipment dispatch, and sales transfer. Human escalation is treated as part of the design surface, not as an external fallback. Builders decide where the agent continues, where a system action occurs, and where a person takes over.

The repeated product terms “control,” “precision,” and “escalation” summarize the positioning. Workflows are meant to make agent behavior more explicit by defining the paths among nodes, subagents, tools, knowledge bases, models, and human operators.

ElevenLabs ties Workflows to confidence at scale through “agent testing integration.” In the interface, testing sits alongside the same configuration surfaces that define a subagent’s knowledge base, tools, voice, and model choice. The reliability claim is not only that an agent can be prompted well, but that more of its operating environment can be designed directly.

That design environment includes the route a caller should take, the subagent responsible for that route, the knowledge and tools available at that step, the model selected for the task, and the conditions under which escalation happens. Verification and authorization are also part of the product framing, along with transfers to human operators.

That is the practical shift for builders. Workflows do not replace prompts; the interface still includes global prompting and node-specific conversation goals. But the control burden is no longer placed on prompting alone. It is distributed across routing, scoping, tool access, knowledge-base selection, model choice, testing, and handoff design.

The final screen directs builders to elevenlabs.io/agents. The product claim attached to Workflows is control: a way to design conversation paths, assign narrower subagents, connect each stage to the right knowledge and tools, and decide when the system should act, answer, verify, or escalate.