70% of PRs Are Submitted by Agents, Yet Costs Drop 52%: Uber's Six Published Default Settings

70% of Uber's PRs are submitted by agents. Weekly active users grew 7x, and cost per session fell 52%.

These two numbers together are a bit counterintuitive. Usually, when usage grows 7x, the bill grows with it. Uber's approach to saving money is not about the model. Their official blog makes it clear: the cost reduction comes from default configuration; the model itself did not become smarter.

Let's lay out this configuration first.

Uber's coding agent runs both locally and in the cloud, and is used by engineers and non-engineers alike. It has more than 3,600 shared skills and runs more than 30,000 times per day. The main model handles planning and checking results, while heavy work such as reading and searching is handed off to cheaper subagents.

The diagram also mentions Context Graph, which provides query interfaces to more than 30 internal systems, with 24 million nodes and 80 million edges. All tools go through a unified MCP gateway for authentication and management.

Specifically, Uber named six default settings as copyable:

1. Compress at 400K tokens; don't wait until it's full

Even if the model supports a 1M context, Uber still performs automatic compression at 400K tokens. The longer the context, the more expensive each call is, and the more scattered the model's attention becomes. Compress early; don't wait until it's packed full.

2. Reasoning effort defaults to Medium

Raise it when needed. Most tasks do not need the highest setting.

3. Let cheap subagents do the heavy work

The main model breaks down tasks and accepts results. Reading files and searching code are handed to cheaper models.

4. Cache in layers by role

Set the main session's prompt cache to 1 hour and the subagent's to 5 minutes. Expensive sessions are worth keeping around longer.

5. Don't install every MCP tool

With more than 100 tools installed, the descriptions (schemas) of those tools alone consume 50,000 to 70,000 tokens before the user has typed the first sentence. Uber's approach is to route tool calls through the CLI and search and load them only when needed. One commenter made an analogy: it's like paying an entry fee before you even get in the door.

6. Combine batch tool calls into one script

This cuts tokens by 55% to 71%, and it works for small queries too.

After this was made public, there were two kinds of reactions in the comments.

One was about maintenance. Dusan pointed out that more than 3,600 shared skills need someone to own them; one outdated skill can push a bad PR to everyone. Quantity itself is not an asset; ownership is.

The other was about validation. Ali Shaheen said the post did not clearly explain how validation and testing are done. His company, palgu.ai, does not assign models by subagent but preconfigures models by task type; for example, Gemini works best for spec validation. This gap is worth noting: the cost-saving part is explained in detail, while the quality-assurance part is left blank.

Others offered more conservative starting points. Myttle said they would first test the 400K compression default. Gipp suggested doing compression first and leaving batch scripts for last.

None of the six requires changing models. What changes are all default values: compression threshold, reasoning effort, which model the subagent uses, how long to cache, how tools are loaded, and how calls are combined. For most teams, this is the lowest-cost step. Pick one you can change today; starting with compression is fine.