Engineering brief
Uber's AI edge isn't agents. It's a decade of platform work.
At a glance
The brief is enough
- Relevance
- Practical value
- Radar
- Time horizon
- This year
- Warnings
- None
Uber built its AI agent stack on a decade of developer platform investment: model gateways, MCP auto-generation, and Slate for isolated testing. The agents are the tip of the iceberg.
AI agent productivity is capped by platform maturity, not model capability. Invest in foundations.
Watch if you're planning a multi-year platform investment and want concrete examples of where to start.
Summary
The core insight. Uber's AI advantage didn't come from clever prompting or the latest model. It came from a decade of building developer platforms: monorepos with Bazel, cloud dev pods (2017), an ML platform (Michelangelo, 2015), distributed tracing (Jaeger), Slate for isolated testing environments, and a unified API gateway. Each of these pre-AI systems became a building block that AI agents could hook into with relatively little friction. The company didn't start from zero in 2023; it started from a decade of disciplined platform engineering.
The model gateway. The first major AI-specific investment was a model gateway in late 2023, built on top of Michelangelo. It exposes a single OpenAI- or Anthropic-compatible API across all vendors and model weights. Every model request at Uber goes through this gateway. On top, Uber added AI Guard, a middleware layer that enforces content safety, security policies, and per-user spend tiers. The result: teams can switch models without changing code, and the platform team can enforce governance at one choke point.
The MCP gateway. In April 2025, Uber built an MCP gateway modeled on the same principle. Because every internal service already publishes thrift or protobuf IDLs to a central registry, the platform team auto-generates MCP server specs for any service. No code changes required. This means agents can call any internal API as an MCP tool, and the same team that built the original API gateway built this one.
The agent workflow. Uber's background coding agent, Minion, handles about 15% of production PRs. It runs on a mega dev pod with all repositories cloned and builds cached. Minion can create PRs across repos, self-validate using Slate (the pre-existing isolated testing environment), and attach screenshots or videos to the PR for human review. The key is that agents don't need new testing infrastructure; they reuse Slate, which took years to build and required every service to support trace-based routing.
The governance layer. Uber tracks every model call, every agent execution, and every PR through unified systems. Code Inbox gives engineers a dashboard of all PRs waiting for them, with AI-generated risk scores and blast radius estimates. The company is building toward a 'green lane' for low-risk PRs to auto-merge, but only after backtesting against years of incident data. Spend tiers require manager approval, and engineers get personalized cost feedback. This is governance as a product, not as a policy document.
My take
I'm impressed by how cleanly Uber retrofitted AI onto existing platforms. But I'm cautious about the lesson most teams will take. It's tempting to hear 'build a model gateway' as the actionable takeaway. The real takeaway is less satisfying: invest in developer platforms for a decade, and then AI agents will be easy. For most organizations, that means picking one platform primitive, like a monorepo or cloud dev environments, and doing it well. Ignore the agent hype until you have the foundation. The agent harness is the easy part.
What to do with this
- Audit your current developer platform: do you have a monorepo with a single build system? If not, start there.
- Evaluate your testing infrastructure: can agents create isolated environments (like Uber's Slate) without human setup?
- Assign one team to build a model gateway that enforces security, spend limits, and observability across all AI calls.
Watch the video
This video is blocked due to your privacy settings. To watch this video, please accept YouTube marketing cookies.
Related breakdowns
Local AI crosses a threshold: what running 180B models on a $7K
A 180B model on a $7K laptop generates games offline. The demo is real. The question for engineering leaders: is local AI production-ready yet, or still a demo?
OpenCode breach: BYOK AI keys are the new attack surface
OpenCode's breach exposed the BYOK trap: stolen API keys fueled a distillation attack. AI keys now need production-grade governance.
OAuth won't work for autonomous AI agents. Here's what will.
OAuth was built for deterministic software, not agents that improvise. Runtime authorization per tool call, with intent context and budgets, is the emerging…
Get TL;DW
Too Long; Didn't Watch.
A concise breakdowns of the AI and devtools videos that actually matter for engineering leaders.
Free. Weekly. No hype.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.