Uber cut coding agent cost per session 52% with four defaults, including compaction at 400K tokens and no preloaded MCP schemas
Uber's fix for runaway coding agent spend was a set of defaults chosen centrally and A/B tested, which is a pattern any team running agents can copy.
What happened
VentureBeat reported on November 19 how Uber brought its AI coding costs under control after its CTO said in April that the company had used up its 2026 AI budget. Earlier this year Bloomberg reported that Uber capped employees at $1,500 a month in token spend per agentic coding tool.
Uber now sets four defaults for interactive coding sessions. Automatic compaction starts at 400,000 tokens, even on models with million-token windows. Reasoning effort is medium, because Uber found recent models performed better there than at high. The prompt cache lives for one hour for the main agent and five minutes for sub-agents. No MCP tool definitions load at startup; agents call command-line tools connected to a central MCP gateway, because preloading 100+ tool definitions would take roughly 50,000 to 70,000 tokens.
Uber A/B tests these settings across employee cohorts. Uber reported that, holding the model constant, cost per 1,000 requests fell almost 34% from its peak, and cost per session fell 52% from its June peak. Uber engineer Uday Kiran Medisetty gave no savings figure for any single default.
My take
The useful part is the method. Uber treated agent configuration as a platform decision with metrics behind it, instead of a per-user preference. Most teams I work with never look at these settings at all, so they pay for a full context window, maximum reasoning and every tool schema on every turn.
The MCP number is the one I'd act on first. If your agents connect to a CRM, a helpdesk and a few internal APIs, the tool definitions alone can cost more than the actual question. Loading them on demand, or routing through one gateway, is cheap to try.
I'd also copy the measurement side. Uber tracks reverts, incidents per change and review times alongside cost, so a cheaper setting that hurts quality shows up. If you change effort levels or compaction without tracking quality, you only know half the result.
The article doesn't say how each default performed on its own, so I'd test them one at a time on your own workloads rather than adopting all four at once.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
