SirenSpec

YAML-first agent orchestration for humans.

Define multi-agent workflows in plain YAML. Run them against any LLM backend.

SirenSpec lets you describe multi-agent pipelines the same way you'd describe them to a colleague — in plain language, structured as YAML. No framework boilerplate, no hidden abstractions. Just agents, nodes, and edges.

version: "0.1"

agents:
  analyst:
    model: "anthropic:claude-haiku-4-5-20251001"
    system: "You are a concise financial analyst."

nodes:
  summarize:
    agent: analyst
    writes: output.summary

input:
  message: "Summarize Q1 earnings for NVDA."

sirenspec run workflow.yaml

That's it. One file. One command. A complete agent run with a structured JSON trace.

Node types — four primitives cover most patterns:

• Agent — single LLM call, writes to a context path
• Swrm — fan-out to N agents in parallel, optionally synthesise
• Factory — spawn one agent per item in a runtime list
• Tool — call an HTTP endpoint or Python callable, no LLM required

Providers — OpenAI, Anthropic, and Ollama out of the box, selected with a simple provider:model URI.

Guardrails — prompt-injection detection and output-length limits at the workflow or per-agent level.

Retry & failure handling — exponential backoff, fallback nodes, default outputs, and per-error-type retry rules.

Template interpolation — reference prior node outputs, environment variables, and runtime inputs directly in prompts with {{ expr }} syntax.

pip install sirenspec

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...

sirenspec run path/to/workflow.yaml

See the full documentation and cookbook examples to go further.

SirenSpec is open source and welcomes contributions. To get started:

git clone https://github.com/TJLSmith0831/sirenspec
cd sirenspec
uv sync --extra dev
source .venv/bin/activate
pytest

Please open an issue before submitting a large PR so we can align on direction.