Show HN: Rebuno - An open-source runtime for production agents

Rebuno, an open-source execution runtime for AI agents, launched on Hacker News. The tool manages execution state and enforcement policies for agents built with different frameworks, running them as HTTP services and dispatching work through signed webhooks.

The runtime keeps execution state in Postgres and requires agents to submit tool and model calls as steps before executing them. For each step, Rebuno evaluates per-agent YAML policies independently of the model's prompt, allowing calls to proceed, be denied, or held for human approval. Policy decisions and step outcomes are recorded in an append-only event log. The system handles interruptions and crashes by restarting agents from the top — steps with recorded outcomes return those outcomes instead of re-executing. Tools marked safe_to_retry run again after a crash; those marked at_most_once fail as indeterminate, leaving the agent to decide how to proceed.

Rebuno ships with Python and TypeScript SDKs and includes examples built with LangChain, CrewAI, Pydantic AI, Vercel AI SDK, and Mastra, per the GitHub repository. The tool is free to self-host and released under the MIT license. A blog post with diagrams and documentation are available.

The launch addresses operational challenges in production agent deployments — principally the need for governance layers that sit outside model prompts and can enforce or override decisions made by LLM-based systems. By separating policy enforcement from agent logic, Rebuno allows teams to define safety rules in configuration rather than prompt engineering, a pattern increasingly relevant as AI agents move from prototypes to customer-facing systems.