{
  "video": {
    "id": "fjF8EKnxKCU",
    "title": "Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind",
    "duration": 1108,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "fjF8EKnxKCU",
    "title": "Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind",
    "one_liner": "Philipp Schmid builds the same GitHub PR-review agent three times — raw Python loop, then ADK framework, then a remote sandboxed Gemini agent with only an AGENTS.md and a bash script — deleting code at every step to show that markdown files and general-purpose tools now do the job orchestration code used to.",
    "summary": "Starting from Simon's definition that \"an LLM agent runs tools in a loop until it achieves a goal,\" Schmid rebuilds one PR-review agent in three generations and shows the source directory disappearing. Version 1 is hand-written Python: the loop, JSON schemas, tool routing, error handling. Version 2 uses Google's ADK, which infers schemas from function signatures but still leaves you owning the Python plumbing and the hosting environment. Version 3 is a single call to the Antigravity remote agent on the Gemini API with an `environment` parameter: a hosted Linux sandbox, an AGENTS.md, and a script that installs the GitHub CLI — no tools defined at all, and the agent still answers off-script questions by reaching for Google Search.",
    "key_points": [
      "The new Gemini Interactions API replaces turn-based user/model conversation history with a flat 'steps' timeline (user input, reasoning, function call, function result) so you no longer 'abuse the user role for passing back data from an environment'.",
      "Demo 1 (raw Python): a run loop, a separate system-instruction file, hand-written JSON schemas and GitHub API calls. Asked anything outside its tools it replies 'I cannot do this'.",
      "Demo 2 (ADK framework): the agent file and JSON schemas go away — the framework builds schemas from Python function signatures and handles the loop, retries and error handling — but asking 'what's the weather in San Francisco' still fails because no tool was defined.",
      "Demo 3 (Antigravity remote agent, launched at Google I/O on the Gemini API): no source directory at all. Just an AGENTS.md telling the agent it has the GitHub CLI, bash and a filesystem, plus a bash script that installs `gh` if missing. The agent explores the sandbox, finds no gh, installs it, then reviews the PR — and answers the weather question via the built-in Google Search tool (~20°C, 2nd of July).",
      "The `environment` parameter gives a hosted isolated cloud sandbox; sources can be a GitHub repo, a GCS bucket or inline files. A network proxy injects credentials on outbound requests so 'the agent never really sees your credential', and you can restrict which domains it reaches (he registered tokens for both the GitHub API and github.com — one for git commands, one for HTTP — with `domain: all` for uncredentialed web access).",
      "The Agents API lets you register a custom ID bundling system instruction, base agent and base environment, then call it exactly like a Gemini model. Server-side session state and automatic context compaction mean multi-turn is just passing the previous interaction ID.",
      "Cited evidence for the trend: Cursor replaced ~12,000 lines of TypeScript git-worktree orchestration with ~200 lines of agent files; Manus refactored its harness five times in six months; LangChain rearchitected open deep research three times in a year; Vercel removed 80% of its tools and got fewer steps, faster responses and better accuracy.",
      "Extending an agent used to mean a new Python function plus a schema plus wiring; now it means writing a SKILL.md (optionally shipping the CLI tool in the environment) — 'we don't need to change our code'. Agents can also write their own files to remember rules and preferences, or externalize context as a hand-off to pick up later."
    ],
    "takeaways": [
      "Stop micromanaging execution paths — give the agent atomic general-purpose tools (bash, filesystem, an installed CLI) and let the model explore and discover the solution instead of enumerating every function.",
      "Extend capability by adding markdown — AGENTS.md for instructions, rules and behaviours, SKILL.md for context and capabilities — rather than adding code paths.",
      "Own what is actually yours: domain instructions, workflows, clean tool definitions, outcome verification, and above all the evals; hand the loop, tool routing, state, compaction and sandbox infrastructure to the platform.",
      "Use harness complexity as a diagnostic: if your harness grows more complex as models improve, you're overengineering it — build to delete, because better models mean more code you can remove.",
      "To try it Monday: open AI Studio to run the Antigravity harness in your own sandbox (or create an API key; a free tier is in progress) and start building files and skills."
    ],
    "topics": [
      "agents",
      "skills",
      "agent-harness",
      "sandboxes",
      "gemini-api",
      "tool-use",
      "context-engineering",
      "evals"
    ],
    "tools": [
      "Gemini Interactions API",
      "Gemini API",
      "Gemini models",
      "ADK (Agent Development Kit)",
      "Antigravity remote agent",
      "Antigravity IDE",
      "Google AI Studio",
      "Google Search tool",
      "GitHub CLI (gh)",
      "GitHub API",
      "Google Cloud Storage (GCS)",
      "Google DeepMind",
      "Cursor",
      "Manus",
      "LangChain",
      "Vercel"
    ],
    "quotes": [
      {
        "text": "we are going to build uh the same agent, the same GitHub PR review agent in three different ways. And we are going to delete code on the way. Each new version, less code, more files basically.",
        "at": "00:36",
        "url": "https://www.youtube.com/watch?v=fjF8EKnxKCU&t=36s"
      },
      {
        "text": "we don't need to change our code. We just provide more files to the agent and the agent decides on what we want to do.",
        "at": "15:16",
        "url": "https://www.youtube.com/watch?v=fjF8EKnxKCU&t=916s"
      },
      {
        "text": "But if your harness is getting more complex as the model improves, you are most likely overengineering your harness.",
        "at": "16:09",
        "url": "https://www.youtube.com/watch?v=fjF8EKnxKCU&t=969s"
      },
      {
        "text": "Agents are just files. We write markdown files to extend capabilities.",
        "at": "16:30",
        "url": "https://www.youtube.com/watch?v=fjF8EKnxKCU&t=990s"
      }
    ],
    "words": 3834
  },
  "summary_url": "/#fjF8EKnxKCU",
  "transcript": {
    "html": "/transcripts/fjF8EKnxKCU.html",
    "txt": "/transcripts/fjF8EKnxKCU.txt",
    "vtt": "/transcripts/fjF8EKnxKCU.vtt"
  }
}