Uber says agents are behind more than 70% of its pull requests, and explains how it cut cost per session 52% while usage grew 7x
Uber's engineering blog describes its 'software factory': interactive harnesses plus managed agents that review code, fix failing CI, complete end-to-end PRs with visual checks, triage on-call alerts and do maintenance, with human review and escalation. More than 70% of PRs are attributed to local or cloud agents, engineers have built over 3,600 agent skills, and skills run more than 30,000 times a day. Most of the post is about cost: benchmarking each managed agent on its real work to pick models, cheaper subagent defaults, compaction at 400K tokens, medium reasoning by default, a one-hour prompt cache, exposing 1,000+ MCP tools as CLI commands, a code graph for context, and a live cost counter in the terminal.
The numbers (as reported)
| Agent-attributed PRs | More than 70% of pull requests |
|---|---|
| Agent skills | Over 3,600 built; more than 30K executions per day |
| Growth Feb to Aug 2026 | Weekly active users 7x, weekly agent requests 9.4x |
| Cost with model held fixed | Cost per 1,000 requests down almost 34% from peak; cost per session down 52% from June peak |
| Code-mode token savings | 55-71% fewer tokens on small SQL queries |
Facts
- Company / subject
- Uber
- Tools
- Claude Code
- Codebase
- brownfield
- Work type
- Company rollout
- Stack
- —
- Published
- 2026-08-27
- Last checked
- 2026-10-27
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- More than 70% of pull requests are attributed to local or cloud agents 70%+
“More than 70% of pull requests are attributed to local or cloud agents.”
- With the model held fixed, cost per 1,000 requests down almost 34% from peak and cost per session down 52% from the June peak -34% / -52%
“cost per session is down 52% from its June peak”
- Code-mode cut tokens by 55% to 71% on small SQL queries 55–71% fewer tokens
Caveats
Company writing about its own platform. 'Attributed to agents' isn't defined in the post. Mostly about cost and tooling rather than one delivered project. A related Pragmatic Engineer interview says the background agent Minion writes about 15% of production PRs, with a person reviewing every PR.
More stories by tool: Claude Code
Related topics: Company rollouts and costBrownfield codebases