Urgent.News

What's breaking now, across thousands of outlets.

AI

How the BMW Group governs hundreds of agents across thousands of developers

When hundreds of agents serve thousands of developers, governance cannot be an afterthought. The BMW Group shares the practices and patterns its R&D teams use to build, govern, and operate AI agents at enterprise scale without sacrificing speed, quality, or compliance.

BMW Group's R&D employs over 12,500 developers who collaborate on approximately 500 million lines of vehicle software code. This software is complemented by the BMW Connected Vehicle backend, a cloud-based system composed of more than 1,300 microservices that deliver digital services for 24 million vehicles in operation. Maintaining stringent quality, stability, and compliance standards is crucial in regulated industries.

While frameworks and guidelines can help create agents, managing and operating them effectively at scale remains challenging. Scaling agentic AI across a large organization necessitates a collaborative approach, local experimentation, and knowledge sharing among teams. This generates a dichotomy between development teams' need for rapid innovation and the enterprise's requirement to enforce safety, security, composability, and cost-efficiency standards.

With hundreds of globally distributed teams developing agents for the BMW Group, striking the right balance is essential.

The solution involves building agents that can cooperate from the outset, using an architecture that enables diversity, governability, and automation without relying on a single framework or toolchain. The agent harness, a runtime layer around a large language model (LLM), serves as the execution loop, tool dispatch, memory management, and observability hook for autonomous operations.

To ensure consistent governance, implement a Model Context Protocol (MCP) gateway, central agent registry, and orchestration boundary. This approach makes governance rules, such as authentication, authorization, rate limiting, and content filtering, enforceable at the integration layer, regardless of framework migrations, model swaps, or team reorganizations.

Governance should be proactive and applied throughout the entire lifecycle, with three key properties: framework-agnostic and externalized, domain-sensitive and graduated, and deterministically enforced. These properties ensure governance rules are consistent, adaptable to various agent actions' risks, and enforced outside the agent's reasoning loop.

Additionally, instrument, verify, and evaluate continuously by capturing observable agent trajectories, using ground-truth datasets for evaluation, and integrating observability into the CI/CD pipeline. This ensures agents maintain policy compliance, high quality, and operational efficiency while retaining the speed and innovation necessary for rapid development.

Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at aws.amazon.com →

More in AI

More from Thursday 19 November →