{
  "video": {
    "id": "Y_p3y6zP5QE",
    "title": "Build Your Own Team of Managed Agents - Jack Wotherspoon, Google DeepMind",
    "duration": 1184,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "Y_p3y6zP5QE",
    "title": "Build Your Own Team of Managed Agents - Jack Wotherspoon, Google DeepMind",
    "one_liner": "Gemini managed agents turn a sandboxed cloud coding agent into a single Interactions API call: you name the agent, set environment to remote, and get a Google Cloud sandbox with the Antigravity harness built in. You can share environment IDs between agents so they build on the same files in multi-agent pipelines.",
    "summary": "Jack Wotherspoon (DevEx engineer, Google DeepMind) argues that agents should run in cloud sandboxes rather than on your laptop. A laptop doesn't scale, and running verification locally steals your browser and throws up pop-ups. Building that yourself (VMs, a harness, networking, firewalls) is tedious. Gemini managed agents in the Gemini API package it as one API call. He walks through the building blocks: the harness, environments, skills, credentials and triggers, plus the separate interaction and environment primitives. He then demos a chain of financial-analysis agents (an SEC filing researcher, a quant agent and a report writer) that share one sandbox, and a ticker-analysis app from the AI Studio gallery.",
    "key_points": [
      "A managed agent is called through the Interactions API, the new agent-first SDK that went GA earlier this year. Instead of a model you name an agent (the base Antigravity agent), set environment=remote, and pass normal input.",
      "The sandbox runs on Google Cloud infrastructure. Built-in server-side tools include Google Search, code execution and file read/write, and you can add custom functions or remote MCP servers. You can upload agents.md files and skills, or load a GitHub repository at startup.",
      "There are five building blocks. The harness is the same Antigravity harness used in Google's Antigravity products. Environments are their own entity. Skills are uploaded playbooks. Credentials inject tokens or MCP OAuth. Triggers run an agent on a schedule, e.g. 8:30am before the market opens, with webhooks out to Slack or a morning summary of a team's Jira board.",
      "Interactions and environments are two separate stateful server-side primitives. Interactions hold the conversation: pass the previous interaction ID to continue to the next turn. Environments hold the sandbox file system: pass an environment ID to a new agent and it gets a fresh context window but the same files.",
      "Secrets never enter the sandbox. A proxy between the sandbox and the API puts a dummy placeholder token in the sandbox and swaps in the real secret (e.g. a GitHub token) only when the outbound call matches the target API, so a rogue agent can't misuse it.",
      "You can switch models with a one-line parameter in the agent config: Gemini Flash for fast, efficient work, Gemini 4 Argon (launched the day before the talk) for deep reasoning. Gemini 4 Argon has 'long decode': instead of the usual ~64K output-token max, the 1M-token window is shared across input, thinking and output, so very long outputs such as financial models are possible.",
      "agents.create saves a refined agent so you can invoke it later by agent ID. Every action is recorded as a step trace (e.g. function result for write-to-file, code execution) for debugging and optimisation.",
      "In the demo, an SEC filing researcher pulled Nvidia metrics into metrics.json. A quant agent reusing the same environment ran Python on those metrics to produce a valuation and share-price forecast. A third agent turned that into a report. Separately, an AI Studio gallery app analysed the VOO ticker by fanning out to several managed agents, read Vanguard documents, and produced a PDF report with charts and source links."
    ],
    "takeaways": [
      "Run verification-heavy agent work (writing and running tests, browsing) in a cloud sandbox instead of on your laptop, and to scale, make more API calls rather than managing infrastructure.",
      "Split broad tasks into narrow, single-purpose agents (research, then analysis, then report) that hand off through files in a shared environment ID. Each gets a fresh context window, and the work stays controllable and measurable.",
      "Port your local setup (agents.md, skills, GitHub repos) into the managed agent's sandbox so your personal coding-agent playbooks become reproducible and shareable, and can be built into apps without your personal subscription.",
      "Give agents credentials through the proxy-injection mechanism rather than placing real tokens in the sandbox.",
      "Choose the model per workflow (Flash for speed, Gemini 4 Argon for deep or long-output reasoning), use triggers for scheduled jobs like daily briefings, and read the step traces to tune workflows."
    ],
    "topics": [
      "agents",
      "managed agents",
      "sandboxes",
      "multi-agent",
      "agent security",
      "scheduling",
      "gemini api"
    ],
    "tools": [
      "Google DeepMind",
      "Gemini API",
      "Gemini managed agents",
      "Interactions API",
      "Antigravity harness",
      "Gemini 4 Argon",
      "Gemini Flash",
      "Google Search",
      "Google Cloud",
      "MCP",
      "agents.md",
      "GitHub",
      "Slack",
      "Jira",
      "Google Docs",
      "Google Drive",
      "AI Studio",
      "Nvidia",
      "Vanguard"
    ],
    "quotes": [
      {
        "text": "And really now it boils down to just being a single API call.",
        "at": "02:51",
        "url": "https://www.youtube.com/watch?v=Y_p3y6zP5QE&t=171s"
      },
      {
        "text": "And the other reason is that your laptop doesn't scale.",
        "at": "01:32",
        "url": "https://www.youtube.com/watch?v=Y_p3y6zP5QE&t=92s"
      },
      {
        "text": "So, another big part of the security aspect of managed agents is that secrets never actually enter the sandbox.",
        "at": "12:21",
        "url": "https://www.youtube.com/watch?v=Y_p3y6zP5QE&t=741s"
      },
      {
        "text": "you can have very kind of niche and specific targeted agents that have one core purpose. They do that purpose really well.",
        "at": "13:55",
        "url": "https://www.youtube.com/watch?v=Y_p3y6zP5QE&t=835s"
      }
    ],
    "words": 4141
  },
  "summary_url": "/#Y_p3y6zP5QE",
  "transcript": {
    "html": "/transcripts/Y_p3y6zP5QE.html",
    "txt": "/transcripts/Y_p3y6zP5QE.txt",
    "vtt": "/transcripts/Y_p3y6zP5QE.vtt"
  }
}