Agents need infra, not scripts. Orloj lets you declare agents, tools, policies, and workflows in YAML and runs them with scheduling, model routing, tool isolation, and policy enforcement. Repo: https://github.com/OrlojHQ/orloj
Architecture: orlojd (API, scheduler, Postgres or in‑memory dev store) + orlojworker (claims tasks, routes model calls to OpenAI/Anthropic/Ollama, executes tools). Isolation options: direct, sandboxed, container, WASM. Reliability: leases, capped retries, idempotent replay,
The piece people skip: runtime governance. AgentPolicy, AgentRole, ToolPermission evaluated inline before every agent turn/tool call. Actions fail closed, return structured errors, and generate audit trails. You can budget tokens per run, whitelist models, and block tools.
Takeaway: treat multi‑agent systems like infra-as-code with runtime gates. For agencies, legal, and finance teams that need human approval and auditability, this is a practical scaffolding—not hype. It’s v0.1.0 (Apache‑2). Which policy would you enforce first?
Top comments (0)