AI Agent Requests Rose 9.4x—How Uber Kept Its Bill Stable

Introduction

For enterprise AI coding, the central challenge is no longer whether a model can produce code. It is whether thousands of agent sessions can run reliably, efficiently, and with predictable cost. Uber’s description of its software factory offers a useful example. From February to mid-August 2026, weekly active users across its agent products grew sevenfold, while weekly agent requests increased 9.4 times. Yet total AI spending had remained broadly stable since April. In a controlled comparison using a fixed model, cost per 1,000 requests fell by nearly 34% from its peak, and cost per session dropped 52% from the June peak.

The important point is that the savings did not come from one cheaper model. They came from removing waste throughout the agent execution loop.

The main techniques

Benchmark the work, then route the model. Uber evaluates models on real software tasks rather than relying only on generic benchmarks. Its comparisons include cost per completed task, quality, reliability, latency, timeouts, noise, and metrics such as precision, recall, and F1. Configurations on the Pareto frontier are preferred. A primary agent can use a stronger model for decomposition and evaluation, while sub-agents handle narrowly defined tasks with a more economical model.

Treat every request as a token budget problem. Each conversational turn may resend the entire history, project context, and tool output. Uber uses automatic compression at 400,000 tokens and sets medium reasoning intensity by default, limiting expensive output and reasoning tokens where higher effort adds little value. Prompt-cache TTLs are also matched to usage patterns: longer-lived interactive sessions benefit from a one-hour cache, while short-lived sub-agents retain a shorter lifetime.

Keep tool definitions out of the default context. Loading large MCP schemas into every session creates an upfront and recurring token burden, even when most tools are never used. Uber routes MCP traffic through a central gateway, exposes tools through CLI commands, and supports tool discovery so that schemas are loaded only when needed. For frequently used workflows, code mode moves polling, loops, and batch operations into a subprocess. The model receives a compact result instead of every intermediate step. Tests with repeated SQL queries showed token reductions of more than 50%, with larger gains for batch workflows.

Replace blind exploration with grounded context. In a large code and data environment, agents can spend more time locating information than producing an answer. Uber’s context graph links services, teams, incidents, pull requests, design documents, datasets, and historical usage records. By querying this graph, agents can start from relevant facts instead of launching repeated searches and speculative sub-agents.

Make waste visible. Runtime cost counters, shared spending tiers, alerts, approval flows, and a session analysis dashboard give developers feedback during and after execution. The dashboard flags patterns such as an unnecessarily strong model, an inflated context window, expired caches, or oversized tool initialization, then connects each pattern with a possible remedy.

Why it matters

Uber’s case suggests that enterprise AI cost control should be managed like an engineering discipline. Teams need a target outcome, a workload-specific evaluation set, and continuous measurement across quality, reliability, latency, and cost. Cutting access to capable models is only one option—and often not the best one. Removing repeated inputs, unnecessary rounds, tool noise, and uninformed exploration can preserve quality while reducing spend.

The figures should not be treated as universal benchmarks. They come from Uber’s workloads, internal infrastructure, and routing decisions based on public pricing. Different repositories, teams, and agent designs may produce different results. What is more transferable is the operating model: treat hosted agents as a production system, optimize the entire runtime, and make every efficiency claim traceable to a real task evaluation.

Source: InfoQ 中文