Jon and Kristiane have launched Orloj, an open-source orchestration runtime for multi-agent AI systems under the Apache 2.0 license. The tool addresses a critical gap in production AI: running agents today resembles container management before Kubernetes, with ad-hoc scripts and no governance. Orloj lets developers declare agents, tools, and policies as YAML manifests that the system schedules, executes, and governs.

Built around the concept of 'agents-as-code,' Orloj treats AI agents like infrastructure resources. Developers define an agent's model, tools, permissions, and execution limits in version-controlled YAML. The system supports DAG-based orchestration with pipeline, hierarchical, and swarm-loop topologies. What sets Orloj apart is its governance layer: AgentPolicy, AgentRole, and ToolPermission resources are evaluated inline during execution, creating runtime gates that fail unauthorized actions with structured errors and audit trails.

The architecture uses a server/worker split with orlojd hosting the API and orlojworker instances executing tasks. Tool isolation runs in containers, WASM sandboxes, or process isolation based on risk level. The system includes lease-based task ownership, capped exponential retry with jitter, and dead-letter handling for reliability. Native MCP support auto-discovers tools from MCP servers while maintaining policy enforcement. With 20 operational workflow templates planned and community contributions welcome, Orloj v0.1.0 is available now on GitHub.