Skip to main content

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 long passage is compressed four times in sequence, each agent targeting half the word count of its input. Watch how meaning is preserved, simplified, or lost across rounds — and observe the length guardrail firing when an agent produces output that exceeds the limit.

​
What it demonstrates

  • The length guardrail applied to every hop in a chain
  • Reusing a single agent definition (summarizer) across four nodes
  • How context accumulates: each hop’s output is visible in the trace as working.round_N
  • Emergent behaviour: meaning collapse is gradual, then sudden

​
Run it

sirenspec run docs/cookbook/compression-gauntlet/workflow.yaml
# Try your own document:
sirenspec run docs/cookbook/compression-gauntlet/workflow.yaml \
  --input "A long passage of your choice (150+ words works best)."
Check working.round_1 through working.round_3 in the trace alongside output.final to see compression at each stage.

​
Workflow

docs/cookbook/compression-gauntlet/workflow.yaml
version: "0.1"

agents:
  summarizer:
    model: "openai:gpt-4o-mini"
    system: |
      You are a ruthless editor. Rewrite the passage at exactly half the word count
      of the input. Preserve the most important facts. Drop examples and qualifiers first.
    guardrails:
      - length

nodes:
  round_1:
    agent: summarizer
    writes: working.round_1
  round_2:
    agent: summarizer
    writes: working.round_2
  round_3:
    agent: summarizer
    writes: working.round_3
  round_4:
    agent: summarizer
    writes: output.final

edges:
  - from: round_1
    to: round_2
  - from: round_2
    to: round_3
  - from: round_3
    to: round_4

guardrails:
  - injection
  - length

​
Graph

​
Next steps

Telephone Game

Five agents relay a message — watch drift without compression.

Guardrails

How the length guardrail works and how to configure it.