BMW Moves AI Agent Rules for 12,500 Developers Into CI/CD and a Shared Registry
BMW Group has 12,500 developers, and hundreds of teams build AI agents in their own way. To avoid a zoo of incompatible agents, the company moved security rules out of the agents' code and into CI/CD (the pipeline that checks every release), an MCP gateway and a registry. An agent drafts documentation almost on its own, but writing to a production database requires human approval. The takeaway: it is cheaper to build checks in from the start than to clean up «AI islands» later.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
BMW Group's R&D organization employs more than 12,500 developers, and hundreds of teams around the world build AI agents: programs based on a language model that call tools and carry out tasks on their own. If every team builds them differently, the result is a zoo that nobody can verify or maintain. That is why BMW does not hard-code security rules into each agent.
Instead, the rules live in three shared places: CI/CD (an automated pipeline that checks every change before it ships to production), an MCP gateway (a single point through which agents reach tools) and a shared agent registry. Consider an agent that drafts documentation and another that updates a production database. Their risk differs, and checking both by the same rules makes little sense.
BMW engineers described the approach on the AWS blog. The practical upshot for readers: this is a way to keep control over dozens of agents without slowing teams down.
Scale and the Core Idea
The context is impressive. BMW's R&D organization maintains about 500 million lines of automotive software code. The Connected Vehicle backend consists of more than 1 300 microservices that deliver digital services to 24 million cars on the road. In a regulated industry, the bar for quality, stability and compliance is very high.
Teams must be free to experiment, but the company has to uphold standards for security, compatibility and cost. Without a balance, you end up with fragmented «AI islands» and duplicated work. BMW does not impose a single framework. Rules for authentication, authorization, rate limits and content filtering work at the integration level, not inside the agent's code. As a result, they survive a change of framework, a swap of model and a team reorganization.
Three Autonomy Levels and Checks Outside the Agent
BMW assigns a trust level to each step, not to the agent as a whole:
- Level 0 (human-in-the-loop): a human confirms every action. New agents and risky operations start here.
- Level 1 (human-on-the-loop): the agent acts on its own, and a human reviews its decisions afterward. This suits tasks of medium risk with clear inputs, outputs and a way to roll back.
- Level 2: the agent works independently within strict policies that ordinary code enforces, not a request in a prompt. This level is reserved for well-understood operations with a proven track record.
A single scenario can include steps at all three levels. Autonomy increases gradually, based on measured quality, not on gut feeling.
Checks run in CI/CD, outside the agent's reasoning loop: access rights, policy-as-code rules and compliance checks block a release before it reaches production. Neither the agent nor its prompt can override them. Separate egress controls — an allowlist of recipients and a scan for secrets in outgoing data — stop an agent that holds confidential data from sending it out through a call that is formally legitimate.
Quality, Building From Blocks and the Registry
BMW builds agents from three layers: MCP tools (a single operation, such as an API call), skills (a set of tools for a domain task, for example code review) and flows (end-to-end processes such as a full pull request review). Each layer is tested separately, and the seams between them are covered by interaction tests. Open standards tie everything together: MCP, Agent Skills, the A2A protocol for agent-to-agent communication and OpenTelemetry for tracing.
Any change to an agent triggers the pipeline: a prompt edit, a new tool, a model swap. Fast automated regression tests run first, followed by an evaluation of answer quality against a reference dataset with thresholds. The order is deliberate: evaluation runs only on a build that has passed the tests. An agent without tracing fails the check even before evaluation.
Before writing code, the team prepares five artifacts: the boundaries of the agent's tasks, an interaction model, a clear description of the tools, a reference dataset and quality criteria. This is how BMW deals with overloaded agents, whose bloated prompts reduce accuracy and increase latency. If an agent outgrows its task, it is split into narrow specialists. They update and scale independently, and one agent's error does not bring down the whole process. A narrow agent often gets by with a smaller model, which lowers running costs. The principle of least privilege follows naturally: an agent has access only to the data it needs.
The central registry of agents, tools and skills serves as a catalog, as an approval point with an audit log and as a point for automated security checks, including for prompt injection. At registration, it also looks for duplicates that other teams already maintain.
Takeaways
Start with a reference dataset and tracing, not with orchestration. Assign autonomy step by step. Keep policies in the pipeline, not in prompts. A caveat: this is the company's own account on the AWS blog, and the available excerpt cuts off at the section on codification via Terraform and CloudFormation. It contains no cost-savings figures, so the effect cannot be assessed. Still, as a checklist for anyone deploying agents in a regulated industry, the scheme is concrete.
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
- Nov 19, 2026Social Department replaces a week of binge-watching with four QuickSearch queries on AWS
- Nov 19, 2026AWS AgentCore Gets Lambda Hooks and Uses 34% Fewer Tokens Than Alternatives
- Nov 19, 2026Skyflow in AgentCore Gateway hides personal data from AI agents' tools
Key Questions Answered
What are the key details on Scale and the Core Idea?
The context is impressive. BMW's R&D organization maintains about 500 million lines of automotive software code. The Connected Vehicle backend consists of more than 1 300 microservices that deliver digital services to 24 million cars on the road. In a regulated industry, the bar for quality, stability and compliance is very high. Teams must be free to experiment, but the company has to uphold standards for security, compatibility and cost.
What are the key details on Three Autonomy Levels and Checks Outside the Agent?
BMW assigns a trust level to each step, not to the agent as a whole: Level 0 (human-in-the-loop): a human confirms every action. New agents and risky operations start here. Level 1 (human-on-the-loop): the agent acts on its own, and a human reviews its decisions afterward. This suits tasks of medium risk with clear inputs, outputs and a way to roll back. Level 2: the agent works independently within strict policies that ordinary code enforces, not a request in a prompt.
What are the key details on Quality, Building From Blocks and the Registry?
BMW builds agents from three layers: MCP tools (a single operation, such as an API call), skills (a set of tools for a domain task, for example code review) and flows (end-to-end processes such as a full pull request review). Each layer is tested separately, and the seams between them are covered by interaction tests. Open standards tie everything together: MCP, Agent Skills, the A2A protocol for agent-to-agent communication and OpenTelemetry for tracing.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.