From the source
- BMW Group's R&D organization has built a governance architecture for running hundreds of agents across more than 12,500 developers, and AWS is detailing the practices behind it.
- The automaker's developers work on roughly 500 million lines of vehicle software code, supported by a Connected Vehicle backend of over 1,300 microservices serving 24 million vehicles, so the bar for quality and compliance is high.
- The approach centers on an agent harness, a runtime layer wrapped around a large language model that supplies the execution loop, tool dispatch, memory management and observability hooks.
- Governance is externalized to an integration layer including a Model Context Protocol gateway and a central agent registry, so rules survive framework migrations and model swaps.
- Trust is graduated per step rather than per agent, from human-in-the-loop approval through human-on-the-loop review to autonomous operation inside policy boundaries enforced as deterministic code in the CI/CD pipeline.
- The post says agents lacking instrumentation fail validation before evaluation runs, and that the supporting infrastructure cannot be retrofitted.






