Documentation Index
Fetch the complete documentation index at: https://docs.sirenspec.dev/llms.txt
Use this file to discover all available pages before exploring further.
A triage agent classifies the user’s intent, then conditional edges activate exactly one of two downstream handlers. The other node is silently skipped. Only the matching branch executes per run.
What it demonstrates
- Conditional edges with
when: expressions
- Writing to a nested
working.* path (working.triage.intent)
- Safe expression evaluation — only
working and output are in scope
- Branching without duplicating agent definitions
Run it
# Triggers handle_refund
sirenspec run docs/cookbook/conditional-pipeline/workflow.yaml --input "I want a refund on my last order"
# Triggers handle_general
sirenspec run docs/cookbook/conditional-pipeline/workflow.yaml --input "What are your store hours?"
Workflow
docs/cookbook/conditional-pipeline/workflow.yaml
version: "0.1"
agents:
triage_agent:
model: "openai:gpt-4o-mini"
system: |
Classify the user's message as either a refund request or a general enquiry.
Reply with ONLY one word — either "refund" or "general" — and nothing else.
refund_handler:
model: "openai:gpt-4o-mini"
system: |
You are a customer-support specialist handling refund requests.
Acknowledge the request warmly, outline the refund process in two or three
sentences, and let the customer know how long it typically takes.
general_handler:
model: "openai:gpt-4o-mini"
system: |
You are a helpful customer-support agent.
Answer the user's question clearly and concisely in two or three sentences.
nodes:
triage:
agent: triage_agent
writes: working.triage.intent
handle_refund:
agent: refund_handler
writes: output.reply
handle_general:
agent: general_handler
writes: output.reply
edges:
- from: triage
to: handle_refund
when: working.triage.intent == "refund"
- from: triage
to: handle_general
when: working.triage.intent == "general"
guardrails:
- injection
- length
How branching works
After triage writes its one-word label to working.triage.intent, the executor evaluates both outgoing when: expressions. Only the matching node runs. If neither expression matches (e.g. the LLM returns an unexpected label), both nodes are skipped and output.reply is never written.
when: expressions run in a sandboxed namespace. Only working and output are available — no built-ins, no imports. Evaluation errors are treated as False and the edge is silently skipped.
Graph
Next steps
Adversarial Pair
Two agents take opposing sides; a judge scores the debate.
YAML Reference
Full edge and when: expression syntax.
