{
  "video": {
    "id": "ojJXgXLUQFc",
    "title": "The SDLC You Build Yourself — Andy Wong, Andrei Bocan & Shane Wolf, Atlassian",
    "duration": 3469,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "ojJXgXLUQFc",
    "title": "The SDLC You Build Yourself — Andy Wong, Andrei Bocan & Shane Wolf, Atlassian",
    "one_liner": "Atlassian argues that the remaining AI gains are to the left and right of code (planning, review, orchestration, operations), not in the coding step itself, and that what makes those agents work is an organisation-wide context engine that searches live code alongside Jira, Confluence, Slack and past decisions.",
    "summary": "Atlassian's DevAI team (Andy Wong, Andrei Bocan, Shane Wolf) say the productivity gains from AI coding tools are tapering. Their bet is agents that cover the rest of the SDLC: the Jira Coding Agent, AI Reviewer with centralised standards, AutoDev for working through backlogs on its own, AI Planner, and AI SRE. All of them are fed by a context engine built on the Teamwork Graph. The planned hands-on workshop was mostly lost to conference Wi-Fi problems, so most of the session was Q&A. It covered context freshness and authority, sandboxing, billing, model routing, AutoDev's work-in-progress (WIP) and cost limits, and why historical decision context matters for agent-generated code.",
    "key_points": [
      "AI Reviewer uses models Atlassian trained in-house to filter out low-value, nitpicky review comments. They report a roughly 40% resolution rate for review comments, against a 45% human baseline. Standards are written in prose/markdown and can be applied across GitHub, GitLab and Bitbucket Cloud, to selected repos or to repos matching a pattern.",
      "Jira Coding Agent (JCA) runs in a proprietary Atlassian-hosted container sandbox with Docker enabled. It checks the acceptance-criteria checklist on a work item, and if the repo explains how to run the app it can open a browser to verify the change. Jira work can also be assigned to Claude, Cursor or Codex, and Jira automations can trigger code generation.",
      "Context engine: graph databases link Jira, Confluence, Bitbucket and third-party tools like Notion and Slack, with agentic search on top. Context is fetched on the fly rather than snapshotted, because cached context goes stale. Conflicts are settled by a human: Confluence's 'verified' badge marks content as authoritative, and recent results are favoured over old ones.",
      "Code context search let Atlassian keep separate product repos instead of moving to a monorepo, which they describe as taking on the debt of building monorepo tooling. Across roughly 15,000 repos, the engine first narrows the search to where an answer probably lives, then digs into the code.",
      "AutoDev picks backlog items it is confident agents can do autonomously and opens PRs for review. Admins can set monthly credits, tickets per user, which repos it runs on, and work-in-progress limits (for example 3–5 PRs per person per week). Underspecified tickets go to a 'needs requirements' state for a human. Review is 'where things go to die', so AutoDev pings reviewers on Slack at opportune moments.",
      "Plans for AutoDev include bringing an existing codebase into line with new standards through small, gradual PRs that cause little churn, and 'organizational steering' for large migrations, such as a Rails upgrade across thousands of services, with humans only coordinating.",
      "Billing uses Atlassian credits, pay as you go, broken down per product and experience, not per task. JCA supports Anthropic and OpenAI models; Gemini was uncertain. Model choice is at the top of the harness, and per-skill or per-subagent model overrides are possible. Smart model routing is not shipped yet but is being worked on in AutoDev.",
      "JCA can break a Jira epic into smaller tasks, and automation rules can then start JCA sessions on each one, giving 'waves' of agents. AI Reviewer added support for skills less than a month before the talk. Standards cannot yet be mapped to deterministic, CI-style pass/fail checks."
    ],
    "takeaways": [
      "Look for AI gains in planning, review, incident response and orchestration rather than adding yet more coding assistants.",
      "Fetch context live from code and tools instead of caching snapshots, and give humans a way to mark sources as authoritative so stale documents don't win.",
      "Feed agents historical decision context (why code was built a certain way), not just current code, so they don't repeat past mistakes.",
      "When running autonomous agents, set WIP limits per person, cost caps per ticket and per week, and a 'needs requirements' gate for vague tickets. Actively nudge reviewers, because review is where autonomous PRs stall.",
      "Write each ticket with an explicit acceptance-criteria checklist, and put run instructions in the repo, so agents can verify their own work."
    ],
    "topics": [
      "agents",
      "sdlc",
      "code-review",
      "context-engineering",
      "orchestration",
      "autonomous-coding",
      "enterprise-search"
    ],
    "tools": [
      "Atlassian",
      "Jira",
      "Confluence",
      "Bitbucket",
      "Jira Coding Agent",
      "Rovo Dev",
      "Rovo Chat",
      "AI Reviewer",
      "AutoDev",
      "AI Planner",
      "AI SRE",
      "AI Context Engine",
      "Teamwork Graph",
      "Teamwork Graph CLI",
      "GitHub",
      "GitLab",
      "Slack",
      "Notion",
      "Claude",
      "Cursor",
      "Codex",
      "Anthropic",
      "OpenAI",
      "Gemini",
      "Docker",
      "VS Code",
      "OpenClaw"
    ],
    "quotes": [
      {
        "text": "opting for a monor repo ends up meaning that you're opting for the debt of having a monor repo",
        "at": "18:47",
        "url": "https://www.youtube.com/watch?v=ojJXgXLUQFc&t=1127s"
      },
      {
        "text": "a conflence document is just a snapshot of a thing that happens once but like the truth is in the code",
        "at": "19:18",
        "url": "https://www.youtube.com/watch?v=ojJXgXLUQFc&t=1158s"
      },
      {
        "text": "the reviewing step is where things go to die",
        "at": "45:45",
        "url": "https://www.youtube.com/watch?v=ojJXgXLUQFc&t=2745s"
      },
      {
        "text": "historical information about why decisions were made are something that's often left out of um agent generated code",
        "at": "49:12",
        "url": "https://www.youtube.com/watch?v=ojJXgXLUQFc&t=2952s"
      }
    ],
    "words": 10898
  },
  "summary_url": "/#ojJXgXLUQFc",
  "transcript": {
    "html": "/transcripts/ojJXgXLUQFc.html",
    "txt": "/transcripts/ojJXgXLUQFc.txt",
    "vtt": "/transcripts/ojJXgXLUQFc.vtt"
  }
}