{
  "video": {
    "id": "17-YSUHo6Lk",
    "title": "Agentic SDLC at Uber — Uday Kiran Medisetty & Adam Huda, Uber",
    "duration": 1106,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "17-YSUHo6Lk",
    "title": "Agentic SDLC at Uber — Uday Kiran Medisetty & Adam Huda, Uber",
    "one_liner": "Uber's platform team shows the six building blocks — model gateway, MCP gateway, agentified devpods, a managed skills marketplace, a context graph, and the Cortana assistant — behind a \"software factory\" where over 70% of PRs now come from local or cloud agents.",
    "summary": "Uday Kiran Medisetty walks through six infrastructure building blocks Uber built to make agentic engineering work at the scale of a few thousand engineers across 12 tech sites, arguing the leverage comes from centralizing model access, tool access, environments and knowledge rather than from any single coding agent. Adam Huda then takes one feature idea — a better rider pickup location outside a busy World Cup stadium — end to end through that stack: ideation in Slack with Cortana, Figma mockups with two A/B variants, code by the Minion cloud agent in a mega devpod, validation shifted into the inner loop, self-healing CI, and enrollment into scheduled maintenance skills. Their framing is that the six-year investment in monorepos and Bazel laid the foundation, and that the emerging bottlenecks are no longer code generation but CI capacity, experiment capacity, and deciding whether a thing should be built at all.",
    "key_points": [
      "More than 70% of Uber's PRs are now written by local or cloud agents, producing twice the lines of code per engineer year-over-year; they've also run 250+ automated migrations covering 9 million lines of code. Six years of monorepo and Bazel work is credited as the foundation that made this possible.",
      "A model gateway fronts everything as one OpenAI/Anthropic-compatible endpoint with middlewares: Spire-based identity/auth, a data anonymizer redacting 20+ PII types, and an AI guard of five specialized safety/policy models — all running under 100ms. Engineers set a project ID on a vanilla client and get attribution per caller, user and team; 800+ projects and 100M+ model requests per day flow through it.",
      "An MCP gateway solves the fact that thousands of internal APIs were not agent-accessible: an automated crawler projects internal APIs into MCPs with one config change, and SaaS MCPs (Google, Slack, Jira) are hosted centrally with token exchange behind one install path.",
      "Token-tax mitigation evolved in three steps: direct MCP → \"Omni MCP\" (one installed MCP that discovers and invokes any MCP in the gateway) → projecting MCPs into a CLI pattern so responses don't consume context, plus an auto-installed \"code mode\" skill that writes Python scripts on the fly for the heaviest token consumers. Net: 1,000+ MCP tools and 40%+ fleetwide token savings.",
      "Devpods were \"agentified\" into pre-provisioned Kubernetes balloon pods with repositories already snapshotted and search indexes already built, so agents start working in seconds. Because engineer roles are blurring, the per-language devpod (Go, Java, Android) gave way to a \"mega devpod\" holding all repositories, which is what autonomous coding agents now use.",
      "Skills were duplicated, hard to discover and often subpar, so Uber built a managed skills marketplace: 2,500 skills, lint checks and automated reviews for baseline quality, one command to discover and install, persona-based auto-install, and now trace/comment collection and continuous evals fed back to skill authors. 20,000+ skill executions per day.",
      "Traces showed agents burning tokens and latency just finding basic context scattered across 20–30 internal systems, so Uber built one context graph — 150 unique node and edge types, 40 million entries — spanning mobile builds, backend, data lake, design docs, Jira and incident bugs. On a sample text-to-SQL question (\"how many mobility trips in India are cash\"), with-vs-without graph showed massive improvement in tokens, turns and latency.",
      "Cortana, the internal assistant, exposes skills, MCPs and the context graph on Slack, CLI and web; employees can attach custom skills and prompts to a team Slack channel so it behaves like a teammate. 300 unique personas created in one month, 20,000+ sessions per day.",
      "In the demo, the Minion cloud coding agent deliberately stops at a draft PR without pushing to CI — good enough for toil, but real end-to-end features need validation first and CI load needs protecting. Validation shifts into the inner loop: static analysis fixes, launching a simulator via a skill to screenshot and compare against Figma specs, and bringing up the service in backend staging to check frontend/backend integration.",
      "Code review is split by loop: a smaller/medium model runs fast in the inner loop, while the outer loop uses a powerful reasoning model with a review skill. Autonomous diffs carry a table on the PR listing every check they passed, including screenshots, to give the human reviewer confidence.",
      "Maintenance runs as a managed loop, not ad-hoc automation: services are enrolled into maintenance skills (e.g. feature-flag cleanup for the losing A/B variant), scheduled on Sunday when CI capacity is free and with the Monday diff volume deliberately capped. Whether those diffs get comments and land or not becomes labeled data to improve the skill, and incident reviews are mined monthly for new maintenance skills."
    ],
    "takeaways": [
      "Centralize model and tool access behind gateways before scaling agents: one compatible endpoint gives you PII redaction, safety guardrails, per-project/user/team attribution, caching, and the audit traces you need for benchmarking and self-improvement loops.",
      "Treat MCP token consumption as an engineering problem with real headroom — aggregating MCPs behind a single discovery MCP and projecting tools into CLI form (so responses stay out of context) is what bought Uber 40%+ fleetwide savings.",
      "Stop letting agents rediscover org context per task. Consolidate ownership, dependencies, patterns, design docs, tickets and incidents into a single graph rather than making each agent stitch together 20–30 systems via separate skills and MCPs.",
      "Manage skills like a product surface — a marketplace with lint checks and automated review for quality, one-command install, persona-based auto-install, and evals feeding back to authors — instead of letting every team grow its own duplicate copies.",
      "Shift validation left before CI: have the cloud agent stop at a draft PR, do visual validation against design specs and frontend/backend integration checks in the inner loop, and attach the resulting evidence table to the PR so human reviewers can trust an autonomous diff.",
      "Run recurring maintenance agents through one managed, scheduled surface with bounded diff volume, not thousands of unbounded loops — and harvest the land/no-land outcomes as training signal."
    ],
    "topics": [
      "agents",
      "developer-tooling",
      "mcp",
      "platform-engineering",
      "context-engineering",
      "skills",
      "code-review",
      "evals"
    ],
    "tools": [
      "Uber",
      "Bazel",
      "Spire",
      "MCP (Model Context Protocol)",
      "Omni MCP",
      "OpenAI",
      "Anthropic",
      "Kubernetes",
      "Cortana (Uber internal AI assistant)",
      "Minion (Uber cloud coding agent)",
      "devpod / mega devpod",
      "Slack",
      "Jira",
      "Google",
      "Figma",
      "Python"
    ],
    "quotes": [
      {
        "text": "over the last year, all of the investments we made in agentic AI have led to more than 70% of our PRs now either by local or cloud agents. And all of this led to twice the number of lines of code per engineer year-over-year.",
        "at": "00:38",
        "url": "https://www.youtube.com/watch?v=17-YSUHo6Lk&t=38s"
      },
      {
        "text": "now we have thousand plus MCP tools and uh just with these optimization efforts we've saved more than 40% fleetwide savings.",
        "at": "05:30",
        "url": "https://www.youtube.com/watch?v=17-YSUHo6Lk&t=330s"
      },
      {
        "text": "the key thing here is that this is actually a managed loop that you go to, right? We don't want thousands of loops being set up across the company without any bounds.",
        "at": "16:25",
        "url": "https://www.youtube.com/watch?v=17-YSUHo6Lk&t=985s"
      },
      {
        "text": "it's not about you know can we build we know we can probably build it now it's more of a question of should we build it",
        "at": "17:45",
        "url": "https://www.youtube.com/watch?v=17-YSUHo6Lk&t=1065s"
      }
    ],
    "words": 3875
  },
  "summary_url": "/#17-YSUHo6Lk",
  "transcript": {
    "html": "/transcripts/17-YSUHo6Lk.html",
    "txt": "/transcripts/17-YSUHo6Lk.txt",
    "vtt": "/transcripts/17-YSUHo6Lk.vtt"
  }
}