When an Uber engineer opens Claude Code for an interactive coding session, automatic compaction kicks in at 400,000 tokens, even for models with million-token context windows. Reasoning effort defaults to medium, not high. No MCP tool definitions load into the context at startup. The engineer didn’t pick any of those settings; Uber picked them.

Those settings are part of a broader effort to control costs after Uber used up its 2026 AI budget in the first few months of the year, according to CTO Praveen Neppalli Naga, who spoke with The Information in April. 

Unexpected AI costs have disrupted projects elsewhere, too. In VentureBeat's October VB Pulse survey for AI infrastructure and compute, 38% of 156 qualified respondents said their organization had delayed, scaled back or canceled a planned AI project in the past year because it cost more than expected.

Weekly agentic requests grew 9.4x from February to August across Uber's agentic offerings, not just Claude Code, according to Uber's August engineering post. At VentureBeat's Agentic Infrastructure Stack Summit on Oct. 28, Uday Kiran Medisetty, a distinguished engineer at Uber, said overall AI spending had remained roughly flat since April even as agentic requests kept rising. He described AIFX, a wrapper that installs and configures Uber's interactive coding tools and manages their settings.

Four defaults Uber sets for every engineer

Uber A/B tests its settings across cohorts of employees. “Since we can do A/B testing with managed settings, it allows us to experiment on these settings so that engineers don’t have to really think about all of this, and we give the same defaults for them,” Medisetty told the summit audience. Uber checks “input and output metrics for different cohorts,” he said.

Medisetty described four ways Uber controls the cost of interactive sessions, from default settings to how agents access tools. He called interactive sessions the least efficient way to run an agent and said Uber wants to move more of that work into managed agents, where it can select models for specific tasks.

  1. Automatic compaction at 400,000 tokens. An agent sends its whole context back to the model on every request, so a bigger context makes every step cost more. Past 700,000 or 800,000 tokens, Medisetty said, cache misses and per-request input tokens “cost a lot.” He called 400,000 “a good medium balance.” Compaction condenses the earlier conversation into a summary so the session can continue.

  2. Effort set to medium. Effort controls how much reasoning a model does before it answers, and more reasoning means more tokens. Models released in the last few months “perform much better at medium effort level,” Medisetty said, and show “some degradation at high.” He did not identify the models or tasks behind that comparison. 

  3. Prompt cache kept for one hour for the main agent and five minutes for sub-agents. Cached prompt text costs less to reuse than to process again, but longer-lived cache entries cost more to write. Engineers often step away from the main agent, called the orchestrator, for longer than five minutes, Uber’s August post says, so a five-minute cache would expire and the next request would pay full price to reload everything. Sub-agents finish their tasks quickly, so five minutes is enough. Uber measures how often requests are served from the cache and tunes these timers to raise that rate. 

  4. No MCP tool definitions loaded at startup. Uber still operates MCP servers, but agents access them through command-line tools connected to its central MCP gateway. Direct MCP integrations can preload descriptions of connected tools into the model's context. With 100-plus tools, those descriptions take up 50,000 to 70,000 tokens before an engineer types a word, Uber says. Instead, agents invoke individual commands as needed.

Medisetty put no savings figure on any single default, and Uber’s August post does not break them out. The post reports results for all of Uber’s optimizations together. Holding the model constant to isolate optimization effects, Uber reported that cost per 1,000 requests fell almost 34% from its peak, while cost per session fell 52% from its June peak. The one cost figure Medisetty gave on stage was for model choice. Benchmarking models and picking the right one cut the cost per cleanup of stale feature flags by 90%, he said.

Uber is also trying to make more of its agent workloads independent of any particular model. Medisetty said the company has shifted its focus from how much traffic runs on open-weight models to how much it can move between models. That's easier with managed agents than with interactive coding tools, where changing models requires changing engineers' behavior. Uber's August post also says it defaults sub-agents to cheaper models for tasks that don't require frontier-level reasoning.

Bigger context windows aren’t always better

Medisetty framed Uber’s 400,000-token cap in terms of cost. Two other summit speakers said the amount of context also changes how well an agent works.

T-Mobile learned that when it tried loading a large amount of material into an agent. “We said, ‘Oh, we’ve finally got a big context window model. Let’s load all of our customer support documentation into that agent,’” T-Mobile CIO Jeff Simon said at the same summit. “That didn’t work at all.” The agent, Simon said, is “going to get confused as it works through the plethora of documentation.”

T-Mobile now keeps “the context as narrow as possible,” he said, routing questions to specialized agents rather than expecting one model to work through the entire support library.

More context can help with some marketing tasks, said Anjul Bhambhri, Adobe’s senior vice president of engineering, at the same summit. She gave the example of a marketer who might use Adobe CX Enterprise Coworker, the company’s agent for marketing teams, to refine a brand’s holiday campaign, a task that may require knowing which offers ran over the past three years. “There’s a lot more context that has to be given because if that is not given, they may get a very generic response,” she said. 

In one case, she said, looking at historical offers revealed that free shipping produced better conversion among loyal customers than a 20% Black Friday discount.

Bhambhri put it as “the right context for the right task for the right person.”

Uber gives agents a map of the company

Uber’s answer to the same problem, how much an agent should see, is to give its agents a map of the company and let them look up the rest.

“What we give the agents is a mental map of how Uber works,” Medisetty told the summit audience. Uber is not “trying to inject all the necessary context up front,” he said.

Uber's context graph holds information about production services, their owners, datasets, historical queries and other internal systems. Agents initially receive the schema, the structure of what the graph contains, and traverse the rest themselves, retrieving relevant documents and source code as needed. 

By Oct. 28, Medisetty put the graph at close to 200 million nodes and edges, compared with 104 million combined — 24 million nodes and 80 million edges — in Uber's August post. Medisetty said Uber's internal evaluations showed better results, lower costs and faster completion times, although he gave no aggregate figures. 

Uber’s August post gives one example, in which the same model received the same prompt: With the graph, it correctly identified a data table in 38 seconds; without the graph, it spent 20 minutes and 9 seconds and incorrectly concluded the data couldn't be queried.

Uber built the approach in a matter of months. AI is a high company-wide priority at the CEO and CTO level, Medisetty said, which made it easier to win support and put the right people on the project.

T-Mobile tried a similar approach and dropped it. It built a knowledge graph for the agents that answer customers on live calls, but the agents had to search a large graph, got lost in it, and the search took time a live call doesn’t have. “We don’t have a lot of time,” Simon said. T-Mobile now uses “a very simple cache based on how customers commonly ask questions,” he said. The cache uses a key-value database to direct agents toward the appropriate information based on how customers phrase their questions. The two cases serve different needs: Uber's agents navigate internal engineering and data systems, whereas T-Mobile's must answer customers in real time.

How Uber measures its return

Uber’s test of whether any of this works, in Medisetty’s words, is to “maximize the throughput while keeping our quality guardrails the same.”

To track whether engineers are producing more, Uber measures the number of design documents, pull requests and deployments. On the guardrail side, Medisetty watches incidents per code change, how often code is reverted, and review and cycle times as agents push more code through.

Code review gets its own check. Uber tracks whether engineers act on each comment its AI reviewer posts. It aims to post only high-signal comments so engineers stop wading through hundreds of them.

Uber also builds trust in its agents gradually. Continuous integration, or CI, runs automated builds and tests on every code change before it merges. When one of those checks fails, Uber’s self-healing CI attempts a fix. 

Uber started conservatively, sending each attempted fix to the engineer in Slack. “Hey, this is our attempted fix. Do you want to accept it?” Medisetty said. Once a type of CI failure showed a good acceptance rate, Uber switched it to auto-accept and patched the pull request directly. Uber also moved simpler CI checks earlier into the development workflow because starting a separate automated fix from scratch was expensive.

Medisetty said Uber has only recently begun seeing a “real trend line” in its ultimate measures: features shipped overall and per engineer. He gave no figures for any of these measures.

Of the 156 respondents to VentureBeat’s October VB Pulse survey on AI infrastructure and compute, 26% said their organization judges whether its AI spending is worth it by “a measured financial return we could defend to our finance team.” Another 44% said they judge it by a measured operational result, such as time saved or volume handled, but not a financial return. Uber's throughput and code-quality measures fall into that operational category. Medisetty did not provide a calculation of Uber's overall financial return on AI spending.

What to ask your own platform team

Uber’s four defaults turn into four questions any platform team can ask of its own coding tools. What is the context limit, and when does automatic compaction kick in? What effort level runs by default? How long does the prompt cache live for each kind of agent? How many tokens of tool schema load before the first prompt?