Show HN: Provena: Open-Source Library for AI Agent Context Governance

Provena is an open-source library for AI agent context governance, introduced on Hacker News under Show HN from the HN AI Agents stream. The maintainers are actively seeking contributors. The project currently has 7 contributors, 11 issues labeled good first issue, and 17 issues labeled help wanted. Those issues are described as well-scoped and include code examples.

The core claim is architectural: existing tools already cover other control surfaces for agents, but not the input side of the context window. Microsoft AGT addresses what agents do. Guardrails AI addresses what agents say. NeMo addresses how agents communicate. Provena targets the remaining gap—the data that flows into the context window before action, speech, or message exchange.

That input layer includes retriever results, tool outputs, and agent messages. In multi-step agent systems, those sources routinely expand the prompt with untrusted or loosely trusted content. Without governance on that path, policy and safety controls applied later (behavior, output, or inter-agent messaging) operate on a context that was never constrained at ingestion.

For engineers and builders, this matters because context assembly is often treated as plumbing rather than a control surface. Retrieval pipelines, tool results, and agent-to-agent messages can inject policy-relevant data without the same review applied to tools, outputs, or communication channels. A library aimed at context governance sits at the point where those streams are merged into the model’s working state.

In competitive terms, the post positions Provena against a stack that already has named options for action governance (Microsoft AGT), utterance control (Guardrails AI), and communication (NeMo). The differentiator is not another output filter or tool-permission layer; it is governance of the ungoverned input path those systems do not cover.

Practical next steps for readers: inspect the Show HN thread and the issue backlog if evaluating adoption or contribution. The 11 good first issue and 17 help wanted labels, plus code examples, are the clearest on-ramps. Watch whether the project hardens concrete policies for retriever results, tool outputs, and agent messages—the three input classes it names as currently ungoverned.