Google DeepMind’s Philipp Schmid builds a finance-oriented AI coworker with Gemini API Managed Agents in this 17-minute AI Engineer demonstration.

Give the agent seven essentials

  1. A brain: a language model to reason through the task.
  2. A computer: a managed sandbox for code, files, and artifacts.
  3. Playbooks: concise instructions and progressively loaded skills.
  4. Tools: built-in search, MCP connections, or application-owned functions.
  5. Credentials: externally managed access to permitted services.
  6. House rules: protected hooks for deterministic checks and approvals.
  7. A schedule: cron triggers for recurring work; webhooks are described as forthcoming.

Keep context organized

  • Bootstrap the environment with files, a repository, or a storage bucket.
  • Keep always-loaded instructions short and split skills into references and executable scripts.
  • Files can hold working artifacts and cross-session lessons; the talk does not benchmark memory quality.

Separate secrets from the sandbox

  • Schmid describes a network proxy that replaces placeholder tokens with real credentials for configured destinations.
  • The model sees dummy values while credential storage supports replacement without readback.
  • This limits direct secret exposure; authorization, approved destinations, and the consequences of legitimate API calls still need controls.
  • Protected hook files let the surrounding software check selected actions and request approval.

Read the demo at its demonstrated scope

  • A browser-based spreadsheet replica triggers a real managed-agent request using a portfolio-audit skill.
  • The presented workflow retrieves financial information, updates the sheet, and can report through Slack.
  • A follow-up recommendation stays advisory in the shown example; no autonomous trade execution is demonstrated.
  • Product capability and security statements come from the vendor presentation, without an independent security test.

“Context is very valuable for us.” — Philipp Schmid, 8:12