# Shannon

A production-oriented multi-agent orchestration framework that runs agents as Temporal workflows with token budgets, WASI sandboxing, policy-based access control, and replayable, step-by-step execution history.

Shannon is a framework for running AI agents in production rather than as demos. Each agent task runs as a Temporal workflow, so any execution can be replayed step by step for debugging, and hard token budgets per task or agent trigger automatic fallback to a cheaper model when a budget is close to exhausted.

Security and observability are built in: agent-run code executes inside a WASI sandbox, Open Policy Agent policies enforce multi-tenant isolation, and real-time event streaming, Prometheus metrics and OpenTelemetry tracing expose what an agent is doing. It works with OpenAI, Anthropic, Google, DeepSeek, xAI and local models served through Ollama.

It installs with a one-command script that pulls Docker images and prompts for API keys, and can be driven through a REST API or a Python SDK. It is written in Go and released under the MIT license.

Website: https://shannon.run