This skill enables designing and building the infrastructure needed to run agents in the cloud instead of on local machines. Hosted agents provide unlimited concurrency, consistent execution environments, and real-time collaboration ("multiplayer").
Mental Model: The value of a hosted agent is the invisibility of the infrastructure. The session must be ready before the user finishes typing their prompt, eliminating the friction of local setup.
🚩 Fragility and Autonomy
- Fragility: High. Sandbox security and state synchronization are critical and prone to concurrency errors.
- Freedom: Medium in the choice of sandboxing provider (Modal, Cloudflare, AWS), but Low in isolation principles.
🚀 When to Activate
- When building coding agents that must run independently of the user's device.
- To design secure execution environments (sandboxes) for AI-generated code.
- When implementing collaborative (multi-user) agent sessions with shared state.
- To scale agent capacity by creating sub-agents in parallel.
🧠 Core Concepts
1. The Spin-Up Challenge
Spinning up a complete development environment (clone, dependencies, build) takes minutes. Users expect seconds.
- Solution: Image Registry Pattern. Pre-build environment images every 30-60 minutes with all the setup ready.
2. Architecture Layers
| Layer | Responsibility | Typical Technology |
|---|---|---|
| Sandbox | Isolated and secure execution. | Modal, E2B, Docker. |
| API/State | Session management and coordination. | Cloudflare DO, SQLite. |
| Client | User interface and input. | Slack, Web, Extensions. |
3. Multiplayer Synchronization
A single state system that synchronizes changes across multiple clients (Web, Slack, VS Code), allowing several humans to collaborate with the same agent on the same branch.
🛠️ Implementation in Gemini Elite Core
Pattern: Predictive Warm-Up
Start spinning up or preparing the sandbox as soon as you detect typing activity in the client, not after the prompt is sent.
Pattern: Self-Spawning Agents
Create tools so that a main agent can trigger new specialist agent sessions for parallel tasks (e.g., one agent researches the frontend while another debugs the backend).
<TOOL_CALL: spawn_agent>
{
"objective": "Analizar la tabla 'users' en la DB de staging",
"context_pointers": ["docs/db_schema.md"]
}
</TOOL_CALL>
🔗 Integrations
multi-agent-patterns: Coordination of sub-agents in hosted environments.tool-design: Design of communication tools between the sandbox and the user.filesystem-context: Use of persistent volumes to maintain state between sessions.
📚 Internal References
references/infrastructure-patterns.md: Detailed patterns for Modal, Durable Objects, and authorship tracking.