Uber: 70% of PRs now come from agents, session costs down 52% at scale
xiaohu · x · 2026-09-02
Uber's engineering team details how it built a software factory at scale.
Key numbers (Feb–mid Aug 2026)
- Weekly active agent users up 7x; weekly agent requests up 9.4x
- Total AI spend stayed flat while per-session costs fell 52% from peak
- 70%+ of pull requests are attributed to local or cloud agents
- 3,600+ agent skills covering the SDLC, executed 30,000+ times daily
- Code review, CI self-healing, end-to-end PRs with visual verification, alert triage, bug debugging, and maintenance increasingly initiated by managed agents, with humans reviewing or stepping in on exceptions
The supporting system
- Unified model gateway for identity, privacy, budget, and audit
- MCP gateway exposing internal APIs to agents
- Pre-started DevPods for isolated, reproducible cross-repo dev environments
- Skill marketplace governing reusable workflows
- Context Graph providing service, team, data, and incident context
- Products like Cortana and Minion orchestrate a pipeline from research and requirements through coding, CI, review, and long-term maintenance
Related event: Uber's AI Software Factory: Agents Handle 70% of PRs(2 posts)→
More from coding & agent
- Developer asks: is LangChain still worth it vs rolling your own agent harness? — curious_vii · 2026-09-03
- Inference Engineering Is Just a Recipe: vLLM/SGLang, Replicas, Cache-Aware Routing — GabGarrett · 2026-09-03
- Developer accidentally built an entire agent factory with Fable 5.1 — 0xkarasy · 2026-09-03
- Microsoft adds Fabric data agents to Foundry agents via Fabric IQ (preview) — adnan_hashmi · 2026-09-03
- Databricks pitches agent-native data infrastructure, Lakebase Postgres at VLDB 2026 — matei_zaharia · 2026-09-03
- doodlestein ships a comprehensive web app review skill after months of debugging — doodlestein · 2026-09-03