Open-source infrastructure for governing agentic systems.
We work on a simple question: when an AI agent acts, who authorised it, and can you prove it afterwards? Most of the conversation around agent safety focuses on what models can do. We focus on what they're permitted to do, at runtime, with evidence.
Our primary project is MIDAS — an authority governance engine that treats decisions as first-class objects. Where policy engines like OPA answer "is this allowed?", MIDAS answers "does this agent have the authority to execute this decision, in this context, right now?" — and produces a signed audit envelope so the answer is verifiable later.
Built in Go with a Postgres backend. Designed to sit alongside existing policy and identity infrastructure rather than replace it.
This Space will host model artefacts, evaluation datasets, and demos related to agent governance — including reference workloads for testing authority boundaries, drift detection benchmarks, and worked examples from the broader MIDAS ecosystem.
AcceptLabs maintains MIDAS as an open-source project under the accept-io GitHub organisation.