Uber

What it is#

The ride-hailing and delivery company. It appears in this wiki as an adopter that publishes its own internals. Five sources in the corpus describe its AI developer stack. Three are first-party: two Uber engineering-blog posts and a podcast interview with Distinguished Engineer Uday Kiran Medisetty, who leads much of the company's agentic-AI work and authored or co-authored all three. The other two are Gergely Orosz's relays of Uber charts and diagrams. No other organization in the corpus is documented from the model layer up to the review queue. Every number is Uber's own, and none is audited.

The podcast account (Uber's AI infra from the inside (with Uday Kiran Medisetty), 2026-10-14) explains why the stack is in-house. In 2015 off-the-shelf software could not handle Uber's scale or did not exist. Its examples are M3DB (about 500 million metrics a second), Cadence (durable multi-day workflows, later the basis of Temporal), Morpheus (feature flags combined with experimentation for multi-sided marketplaces) and uOwn (service ownership). Cities are a primitive all the way down: each is pinned to a cloud region and can fail over. The service count went from two in 2015 to thousands by 2019, and Medisetty expects about the same number in 2026 against a business roughly 10× larger.

The AI stack, layer by layer#

Medisetty's summary of why the stack arrived quickly: "All the things that we built to make life easier for engineers: it makes it easier for our agents as well!" Orosz frames it as roughly a year of AI tooling built on top of a decade of developer-tooling investment. In this corpus Uber is the strongest case that an AI-ready organization is mostly an already-instrumented one. It is also a case only a company that already had a platform organization of this size can make.

Spend governance#

Three mechanisms, from two first-party sources:

• Spend tiers with manager override (podcast). Each engineer is allocated a monthly spend tier. Past it, their manager can sign off on more. "For the right type of work, there is no spend limit, but Uber doesn't want engineers to carelessly burn tokens." It is a ceiling with an escape hatch, the mirror image of the usage floor measured on AI Usage Mandates and Quota Gaming.
• CostCheck (podcast), a plugin that works across sessions and suggests how an engineer can improve their cost profile. Together with the blog's live per-user cost counters, Slack nudges at 50/80/100% of expected spend and a 16-category anti-pattern dashboard (Running a Software Factory Efficiently at Uber Scale), it is visibility handed to the individual rather than reported upward.
• The cost line. The podcast repeats "reduced token costs by about 50% the last few months" as previously reported. The blog's own figures are cost per 1,000 requests −34% and cost per session −52%, each measured at its trough with one model held fixed. Orosz's chart caption says "per-token costs are down 50%+". These are three different units, so they do not corroborate one another. See Cost-per-Task Over Cost-per-Token and AI Product Economics Maturation.

Throughput, review and CI#

• Code review is the admitted bottleneck. Every PR is still reviewed by an engineer. The tooling exists so the reviewer can decide how much time a change needs. AI review, attached media and a risk score are all in place before a human looks. The risk score here allocates attention. It is not the auto-approval gate on Risk-Tiered Auto-Approval.
• CI code red (early 2026). CI jobs more than doubled in 12 months, and Uber declared an internal code red. The response was to move simple jobs (linting, build and template rules) into the agent's inner loop: agents run afx validate before opening a PR. That is volume outrunning a verification control, the system-level strain on Acceleration Whiplash, plus a fix that moves a check earlier rather than adding capacity. See Layered Supervision and the CI-capex question on Verification as the New Bottleneck.
• PRs tagged by work type, not PR velocity. Uber used to measure how long a diff took to reach production. Agents make PRs cheap and many have no business impact, so every PR is now tagged as feature, bugfix, maintenance, testing or docs. The feature share is rising and the maintenance share falling. Spotify reports the same instrument and the same maintenance direction (Acceleration Whiplash). Neither publishes its classifier.
• >70% of PRs agent-attributed, with 3,600+ agent skills and 30K+ skill executions a day (blog, August 2026).

The engineer's job, in Medisetty's words#

"The job of engineering is, in some sense: 'how do you build the right context for the agents to produce reliable output?'" He names three skills that now matter more: context engineering, multitasking across parallel work, and reviewing and verifying agent output. See Compute Allocator.

Evidence weight#

case-study throughout. The podcast source is a newsletter author's written summary of an interview. The transcript is not in the vault, so nothing attributed to the podcast here is a verbatim transcript reading except the two quoted passages Orosz prints. Uber publishes no classifier for its PR tagging, no CI-job or cost baselines, and no outcome measures for review risk scoring. Directions are the defensible part.

Connections#

• Standardize the Infrastructure, Not the Tools: Uber is the second named org (after Shopify) to put every model call behind one gateway, and the only one that also standardizes the layer above the tools, carrying settings across harnesses with AIFX
• Cost-per-Task Over Cost-per-Token: home of Uber's six-term cost equation and its lever list
• Tool-Catalog Context Cost: Omni MCP and Code Mode, Uber's two routes around the tool-schema tax
• Deterministic Engineering for Agent Code Review: uReview's filter-only pipeline
• Acceleration Whiplash: the CI code red and PR-type mix, read beside Spotify's ledger

Sources#

• Uber's AI infra from the inside (with Uday Kiran Medisetty): Gergely Orosz with Uday Kiran Medisetty, Uber's AI infra from the inside, The Pragmatic Engineer podcast page, 2026-10-14, case-study (first-party interview, summarized by the newsletter author; the transcript is not in the raw). Cited for the 11 takeaways and the four-tool screenshot collage (viewed)
• Running a Software Factory Efficiently at Uber Scale: Medisetty, Uber engineering blog, 2026-08-27, case-study. Cost equation, spend visibility, context graph, agent-attribution figures
• Designing MCP Gateway: Uber's MCP Management Platform: Srivastava et al. (incl. Medisetty), Uber engineering blog, 2026-10-01, case-study. MCP Gateway architecture
• What is happening with code reviews?: Orosz, 2026-09-08, practitioner-opinion. uReview pipeline diagram
• The state of the tech industry in 2026: Orosz, 2026-10-06, practitioner-opinion. The usage-versus-cost chart and its "per-token costs are down 50%+" caption