The core loop
Key components
Workflows (Agents) The conversation logic. A workflow is a graph of nodes (conversation steps) connected by edges (conditional transitions). You define what the agent says, when it moves on, and what data it collects. Runs Every execution of a workflow creates a run. The run record holds the transcript, recording, extracted data, and cost information. Telephony The phone infrastructure. Bananaflow connects to your telephony provider (Twilio, Vonage, etc.) to place and receive calls. The audio streams between the caller and Bananaflow in real time. Transcriber (STT) Converts the caller’s speech to text in real time. The transcript drives both the LLM and the final run record. LLM Processes the transcript and the active node’s prompt to generate the agent’s next response. It also evaluates edge conditions to decide when to move the conversation forward. Voice Synthesizer (TTS) Converts the LLM’s text response to audio and streams it back to the caller. You pick a voice for each agent, and Bananaflow manages the transcriber, LLM, and speech models behind it.How it fits together
When you trigger a call:- Bananaflow instructs your telephony provider to dial the number
- When the caller answers, a real-time audio pipeline opens
- The caller’s speech is transcribed to text
- The transcript is sent to the LLM with the active node’s prompt and conversation history
- The LLM responds – the response is synthesized to audio and streamed to the caller
- When an edge condition is met, Bananaflow transitions to the next node
- When an end node is reached, the call ends
- Post-call: context is extracted, webhooks fire, the run record is saved
Next steps
Workflows & Agents
How the conversation graph works
Calls & Runs
The lifecycle of a call
Context & Variables
How data flows through a conversation
Campaigns
Running agents at scale