BMW Group sets governance model for hundreds of AI agents

BMW Group is developing an enterprise governance model for hundreds of AI agents used across its R&D organization, where more than 12,500 developers work on approximately 500 million lines of vehicle software code. The approach is designed to let teams experiment locally while enforcing standards for safety, security, quality and cost-efficiency.
- BMW Group’s R&D organization has more than 12,500 developers.
- Its vehicle software spans approximately 500 million lines of code.
- Agent trust is assigned per workflow step across three autonomy levels.
- CI/CD pipelines run verification and evaluation before production.
- The architecture separates MCP tools, skills and end-to-end flows.
Governance for a large software estate
In a blog post, AWS describes the governance practices BMW Group uses to build and operate agents at scale. BMW’s software environment also includes the BMW Connected Vehicle backend, a cloud-based system of more than 1,300 microservices providing digital services for 24 million vehicles in the field.
AWS says the scale and regulated nature of this environment make software quality, stability and governance compliance especially important. Hundreds of globally distributed teams are building agents, creating a risk of fragmented “AI islands” and duplicated work if development is not coordinated.
BMW’s stated answer is an architecture that allows agents to cooperate without requiring every team to use a single framework or toolchain. At its center is an agent harness, a runtime layer wrapped around a large language model. It handles the execution loop, tool dispatch, memory management and observability hooks needed for autonomous operation.
Trust levels are assigned to individual steps
AWS says governance should be structural and maintained throughout an agent’s lifecycle. Rules such as authentication, authorization, rate limiting and content filtering are enforced at the integration layer, including through a Model Context Protocol gateway, a central agent registry and an orchestration boundary.
BMW’s model uses graduated trust rather than treating every action in an agent’s workflow alike. Level 0 requires explicit human approval for every action and is intended for new agents and high-risk operations. At Level 1, an agent acts autonomously but presents decisions for asynchronous human review. Level 2 allows autonomous operation within strictly defined policy boundaries for low-risk, well-understood tasks.
The levels are assigned per step, not per agent. A single workflow can therefore combine human approval, human review and autonomous actions. AWS says teams can increase autonomy gradually as confidence grows, based on evaluation data rather than assumptions. Permission checks, policy-as-code rules and compliance validations run in the CI/CD pipeline before an agent reaches production, while egress controls can use recipient allowlisting and outbound secret scanning to reduce the risk of confidential data disclosure.
Evaluation and composition are built into delivery
The framework treats observability as more than checking whether a service returned an HTTP 200 status. Agents must produce outputs that are faithful and policy-compliant, with execution traces that can demonstrate how those outputs were produced. BMW’s approach uses a ground-truth dataset covering common queries and edge cases, established at the start rather than added after deployment.
Intent tracing records an agent’s trajectory, including reasoning steps, tool invocations and context retrievals. OpenTelemetry spans provide a queryable lineage from a user request to the final output. Changes such as prompt edits, tool additions and model swaps trigger CI/CD checks. Automated regression verification runs first, followed by evaluation of non-deterministic output against quality thresholds.
AWS also describes a move from monolithic, generalist agents toward collaborating specialized agents. Teams are expected to define scope, interaction models, tool definitions, a ground-truth dataset and evaluation criteria before writing orchestration logic. The resulting architecture has three tiers: MCP tools as atomic integration contracts, skills as domain-specific compositions of tools, and flows as end-to-end business processes. Each tier is tested independently, while the overall model is connected through open standards including MCP, Agent Skills and OpenTelemetry.


