OSymandias

Most multi-agent frameworks hand you a Python API and leave the infrastructure problem — the queue, the vector store, the scheduler — entirely to you. OSymandias takes the opposite position: `pip install osymandias`, run `osy init` and `osy serve`, and the full stack (PostgreSQL, Redis, RabbitMQ, Qdrant, four Celery workers, and a dashboard at localhost:47759) starts from a single Docker Compose file. From there, the Python SDK lets you define agents, register custom tools, and describe multi-agent plans that the runtime decomposes into scheduled jobs across a DAG. The runtime handles dependency ordering, retries, and state passing between steps.

The differentiating bet here is OS-inspired primitives. Rather than treating agents as one-shot LLM calls, OSymandias borrows concepts from operating systems — job scheduling, process isolation, shared memory — and applies them to agent execution. That architectural decision means agents share a memory layer backed by Qdrant vector search, and the scheduler can coordinate parallel agents without each one managing its own state. The real-time observability dashboard surfaces what each agent is doing mid-run, which changes how you debug failures: instead of reading logs after the fact, you watch the DAG execute.

This fits a Python team that already knows how to run Docker Compose in production and wants an agent runtime that doesn’t require a paid SaaS subscription. Research pipelines, internal writing or analysis workflows, and projects that need to integrate existing agent frameworks into a shared memory and scheduling layer are the scenarios the docs describe. The hard wall is operational: you are running five backing services, and when Qdrant runs out of memory or RabbitMQ queues back up under load, there is no vendor support line — just your team and the GitHub issues tab. Teams that hit that wall and lack infra bandwidth switch to a managed platform rather than debug distributed systems on top of their agent work.

LiteLLM sits in the call path for LLM routing, which means provider switching — say, from OpenAI to a local Ollama instance — is a config change rather than a code change. The MIT license means you can fork, modify, and embed the runtime in a commercial product without restriction. The SDK is distributed via pip, and the docs describe both English and Brazilian Portuguese documentation, suggesting the project has at least two active language communities.