Build your agent

Patterns

Workflow shapes that work well in production, with the JSON for each.

Fast, natural chat

Listen → AI reply → back to Listen. The AI speaks as it writes, so the caller hears the first words soonest. The AI can still end the call by returning hangup.

1"ask_for_input": { "type": "input", "parameters": { "input_variables": { "user_input": "str" } }, "next": "chat" },
2"chat": {
3  "type": "llm-streaming",
4  "parameters": { "system_prompt": "…", "service": "groq", "model": "openai/gpt-oss-120b", "history_key": "conversation_history" },
5  "next": "ask_for_input"
6}

Instant acknowledgement

Run a tiny model and the main model side by side. The tiny one says “Sure, one moment” almost at once; the main answer follows. Give each its own history_key so they don't mix memories, and a stream_timeout on the main one so a very late answer is dropped. The Quick ack + main LLM template builds this.

1"ask_for_input": { "type": "input", "parameters": { "input_variables": { "user_input": "str" } }, "next": ["ack", "main"] },
2"ack":  { "type": "llm-streaming", "parameters": { "model": "openai/gpt-oss-20b",  "history_key": "conversation_history_ack", "system_prompt": "Reply with a two to four word acknowledgement only." } },
3"main": { "type": "llm-streaming", "parameters": { "model": "openai/gpt-oss-120b", "history_key": "conversation_history", "stream_timeout": 8, "system_prompt": "…" }, "next": "ask_for_input" }

Decide, then branch

An AI decision fills in fields; a Say speaks the reply; a Branch picks the path from a field.

1"decide": {
2  "type": "llm",
3  "parameters": {
4    "system_prompt": "Work out what the caller wants.",
5    "llm_return_type": {
6      "speak":  { "type": "str", "description": "What to say." },
7      "intent": { "type": "str", "description": "book, cancel or other" }
8    }
9  },
10  "next": "say_it"
11},
12"say_it": { "type": "out", "parameters": { "variables": ["speak"] }, "next": "route" },
13"route": {
14  "type": "conditional",
15  "parameters": {
16    "input_variables": { "intent": "str" },
17    "mappings": { "book": "booking", "cancel": "cancelling", "default": "ask_for_input" }
18  }
19}

End the call

Send hangup: true with the last line. The line is spoken first, then the call ends:

1"goodbye": { "type": "out", "parameters": { "out_dict": { "speak": "Thanks for calling. Goodbye!", "hangup": true } } }

To let the AI decide, include hangup in its decided fields and pass it on with "variables": ["speak", "hangup"].

Show what the caller said

A Say step with only "variables": ["user_input"], run beside the AI step, puts the caller's words in the transcript without speaking them back. The templates call it transcription.

Next steps

Import, export and code: reuse a workflow elsewhere.
Steps: every setting of every step.

Patterns

#Fast, natural chat

#Instant acknowledgement

#Decide, then branch

#End the call

#Show what the caller said