Map conversation steps on a canvas. Make decisions and fallback routes explicit. Connect conditions to actions and handoffs. Carry only the variables needed by the next step.
For tasks with multiple outcomes.
How it works +
Define the entry point, required caller details, and the outcome each branch should reach.
Connect conditions to actions and handoffs. Carry only the variables needed by the next step.
Test missing inputs, changed intent, and failed actions; give every branch a clear exit.
Single-prompt agents
Describe a focused role in plain language. Get a simple first agent defined. Configure the voice, knowledge sources, and permitted tools for that specific task.
For a narrow, straightforward task.
How it works +
Set the role, tone, allowed scope, and details the agent must collect in its instructions.
Configure the voice, knowledge sources, and permitted tools for that specific task.
Try ambiguous and out-of-scope requests, then refine the instructions against the resulting transcripts.
Multi-prompt agents
Link prompts for different steps or intents. Keep complex behavior organized. Define routing conditions and the context passed between prompts, including collected identifiers.
For journeys with distinct responsibilities.
How it works +
Split the journey into focused prompts with distinct responsibilities.
Define routing conditions and the context passed between prompts, including collected identifiers.
Test intent changes, repeated questions, and transitions back to an earlier step.
Knowledge sources
Connect business information to the agent. Ground answers in relevant material. Organize and update the material so retrieval can find the relevant facts for a caller’s question.
For policies, services and product questions.
How it works +
Choose the documents, FAQs, or pages that contain the approved answers.
Organize and update the material so retrieval can find the relevant facts for a caller’s question.
Test missing and conflicting answers; define when to ask for clarification or transfer to a person.
Actions & templates
Reuse common starting points and functions. Avoid rebuilding every task from scratch. Map action inputs, required fields, credentials, and permitted operations for each connected tool.
For support, scheduling and repeatable processes.
How it works +
Start with a relevant template and adapt its instructions to your process.
Map action inputs, required fields, credentials, and permitted operations for each connected tool.
Validate returned results and failure paths before using the action in a live conversation.
Versions & collaboration
Manage changes and team responsibilities. Make updates easier to review and recover. Assign editing and release responsibilities, then review prompt and tool changes together.
For teams maintaining a live agent.
How it works +
Keep a known working configuration as the reference for new changes.
Assign editing and release responsibilities, then review prompt and tool changes together.
Run the same scenario set before release and retain a version you can return to.
Build around the work you already do.
Reception
Collect a request and route it to the right destination.
Support
Answer from business knowledge and define escalation.
Scheduling
Collect the right details before invoking a booking action.
Shape the agent. Keep the logic clear.
Start with a clear task and a preferred building approach. Describe how success should look.
Bring one clearly defined call outcome, approved knowledge, required data fields, and a handoff destination. Identify the systems the agent may read or update and who will approve the workflow before it goes live.
When should we use a prompt versus a structured call flow?+
A prompt is useful for flexible conversations within clear boundaries. A structured flow suits mandatory steps such as verification, collecting required fields, or obtaining confirmation before an action. Choose based on the risk and predictability of the task.
How do we keep the agent from making unsupported promises?+
Specify what the agent may say, what it must verify in a connected system, and which questions require a person. Test missing knowledge, conflicting records, and callers asking for exceptions.
How much engineering work will our workflow need?+
An informational agent and an agent that updates business records have different needs. Review available connectors, API authentication, required actions, and error handling with your technical team before estimating implementation effort.
What should be signed off before the first release?+
Agree on the conversation scope, knowledge owner, voice and greeting, system permissions, escalation rules, and acceptance tests. Start with a bounded pilot and review outcomes before adding more tasks.
