Google's managed agents: convenience vs lock-in for agent deployment.

The proposal. Google DeepMind's managed agents run coding agents entirely in cloud sandboxes, accessed through a single Interactions API call. Jack Wotherspoon argues this eliminates local infrastructure burden and enables scaling, team sharing, and application integration. The harness is the same anti-gravity system used internally at Google.

Why now. Teams building agent workflows face isolation, security, and scaling problems. Local agents steal browser focus, cannot be shared, and lack reproducible environments. Managed agents address this by providing a sandbox, built-in tools (code execution, search, file I/O), and infrastructure management. The triggers feature allows scheduling agents for daily tasks like market briefings.

The architecture. The system splits state into interactions (conversation turns) and environments (file system state). Different agents can reuse the same environment, enabling multi-agent pipelines where one agent writes a file and another reads it. Skills (agents.md, playbooks) can be uploaded, and credentials are injected via a proxy that never exposes secrets inside the sandbox. This proxy design is a notable security improvement over local agent setup.

Multi-agent workflows. The demo shows a three-agent pipeline: an SEC filing researcher gathers metrics, a quant agent analyzes them, and a reporter generates a PDF. Each agent gets a fresh context window but shares the same environment. This modular approach lets teams compose specialized agents without monolithic prompts. The video also shows a web app that uses multiple managed agents behind a single interface.

The trade-offs. Managed agents trade control and customization for convenience. Teams lose access to custom harnesses, networking configurations, and direct container management. Vendor lock-in to Google Cloud is a real concern. Pricing and latency at scale are not discussed. The security proxy, while clever, adds a layer of complexity and potential debugging difficulty. The demo is impressive but no production usage data or benchmarks are provided.

What's missing. The talk is a vendor pitch supported by live demos, not production experience. No cost comparison with self-hosted alternatives, no reliability metrics, no discussion of failure modes or fallback strategies. Teams evaluating managed agents should run their own pilots and measure cost, latency, and error rates before committing to the platform. The idea is promising but the evidence is shallow.