{
  "video": {
    "id": "qJtcEDc_WJU",
    "title": "Scheduled Agents, Dreaming and Q&A: Claude Managed Agents (Part 3 of 4) — Anthropic",
    "duration": 1591,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "qJtcEDc_WJU",
    "title": "Scheduled Agents, Dreaming and Q&A: Claude Managed Agents (Part 3 of 4) — Anthropic",
    "one_liner": "Claude Managed Agents treats a single session as the unit of work. Deployments run that session on a schedule, and 'dreaming', in research preview, reviews the last 30 days of session transcripts to repair a memory store; the speakers say to check every such change with your own evals.",
    "summary": "This is the end of the Claude Managed Agents (CMA) demo plus the Q&A. It first shows a main agent that uses a custom 'dispatch analysis' tool to start one session per stock ticker (Google, Microsoft). It then shows deployments, which run an agent definition on a schedule with initial messages and attached resources. It closes with a preview of 'dreaming': sub-agents read past session transcripts, find gaps and update the memory store. In the Q&A the team positions CMA as an API-first managed tool loop rather than Claude Code in a box. They say multi-tenancy should use workspaces rather than bring-your-own-key, RBAC is on the roadmap, there are no scaling knobs, and evals are how you build confidence in your agent.",
    "key_points": [
      "Multi-session orchestration demo: a main agent checks whether it has enough information, then calls a custom 'dispatch analysis' tool. That starts a separate session for each ticker. Each session runs multiple agents with outcome definitions (self-supervision) and memory stores they learn from.",
      "A deployment takes an agent definition (e.g. a 'research desk weekly memo' agent), a schedule, initial messages and optional resources (memory stores, GitHub repos, files). It creates a new session on each run, e.g. 'research NVIDIA once every quarter'. Runs can also be triggered from the console or from a custom UI built on the API.",
      "Dreaming (research preview): you pick a memory store, a model, optional custom instructions and a time window (e.g. the last 30 days). It starts a managed-agents session whose sub-agents analyse the transcripts of every session that used that memory store. In the demo it found gaps in a skill and updated the memory store to extend it. The plan is for it to also improve skills, agent prompts and sub-agent prompts directly.",
      "Observability: the console's debug view shows every event in a session, including model-call start and finish times, so you can derive latency. There is no built-in OTEL export, but because everything is API calls 'you could build that connector fairly easily'.",
      "Compaction is managed for you so sessions can run for days. They have seen 10-hour sessions in production. Memory stores are independent entities that you attach to whichever sessions you choose.",
      "In CMA, 'harness' means tool-loop execution: managing context and session state. The single session is the primitive that builders compose into their own multi-agent topologies. There is no graph or topology workflow UI yet ('maybe soon'). The team prefers letting Claude choose its own sub-agent topology because it 'ends up outperforming' human-constrained designs.",
      "Multi-tenancy and access: workspaces are the current primitive, and everyone in a workspace can access everything in it. RBAC is on the roadmap. The team advises against bring-your-own-key for multi-tenant apps because you lose observability. There are no scaling parameters, only rate limits. CMA is Claude models only, is not on Bedrock but is on Claude Platform for AWS, and on-prem is being 'seriously considered'.",
      "Inspired by but not the same as Claude Code: Claude Code is a terminal product tuned for coding, while CMA is an API and infrastructure product for bringing your own context and capabilities. Notion's external agents and Atlassian's Jira integrations are built on it in production. Sonnet 5 is supported."
    ],
    "takeaways": [
      "Use a single session as your unit of work. Build fan-out yourself, e.g. with a custom tool that starts one session per item, and use deployments for recurring jobs instead of your own cron infrastructure.",
      "Keep your own eval suite and run it whenever models change or dreaming rewrites a memory store. The team's own answer to 'how do I know it works' was 'the boring answer is just run evals'. Use outcomes (a rubric plus max iterations) to steer agents toward a defined result.",
      "For multi-tenant SaaS, use workspaces (one per customer, or one shared workspace with your own agent-to-customer mapping) instead of bring-your-own-key. Until RBAC ships, add your own security layer on top.",
      "Don't over-constrain sub-agent topology. Let Claude decide, and expect skills and harness scaffolding to get 'thinner and thinner' as models improve.",
      "To move non-engineers from Cowork or Claude Code to managed agents, use the console quick start or the managed-agents skill in Claude Code, which walks you through building a first agent in natural language."
    ],
    "topics": [
      "agents",
      "multi-agent orchestration",
      "scheduled agents",
      "memory",
      "self-improvement",
      "evals",
      "multi-tenancy",
      "observability"
    ],
    "tools": [
      "Claude Managed Agents",
      "Claude Console",
      "Claude Code",
      "Claude Cowork",
      "Claude Platform for AWS",
      "Amazon Bedrock",
      "Sonnet 5",
      "OpenTelemetry (OTEL)",
      "Notion",
      "Atlassian Jira",
      "MCP",
      "Kubernetes HPA",
      "GitHub"
    ],
    "quotes": [
      {
        "text": "We've seen 10-hour sessions in production uh from some of our users.",
        "at": "06:19",
        "url": "https://www.youtube.com/watch?v=qJtcEDc_WJU&t=379s"
      },
      {
        "text": "As cloud gets smarter we have continuously remove those guard rails and let claude just let it cook",
        "at": "07:05",
        "url": "https://www.youtube.com/watch?v=qJtcEDc_WJU&t=425s"
      },
      {
        "text": "The boring answer is just run evals and and evaluate these things.",
        "at": "12:17",
        "url": "https://www.youtube.com/watch?v=qJtcEDc_WJU&t=737s"
      },
      {
        "text": "we have seen that the best model for multi-tenency is not bring your own key because you lose observability",
        "at": "18:52",
        "url": "https://www.youtube.com/watch?v=qJtcEDc_WJU&t=1132s"
      }
    ],
    "words": 5691
  },
  "summary_url": "/#qJtcEDc_WJU",
  "transcript": {
    "html": "/transcripts/qJtcEDc_WJU.html",
    "txt": "/transcripts/qJtcEDc_WJU.txt",
    "vtt": "/transcripts/qJtcEDc_WJU.vtt"
  }
}