This skill enables designing and coordinating systems composed of multiple instances of language models (agents), each with its own context window. Multi-agent architecture is the definitive solution for tasks whose complexity exceeds the reasoning capacity of a single saturated context.
Mental Model: Sub-agents are not just roles ("researcher", "writer"); they are context isolation units. Their main purpose is to partition knowledge to maintain a high signal density at each step of the process.
🚩 Fragility and Autonomy
- Fragility: High. Poorly designed coordination introduces massive latency and information loss ("telephone game" effect).
- Freedom: Very High in choosing the orchestration pattern according to the nature of the task.
🚀 When to Activate
- When a task requires handling multiple technical domains simultaneously.
- To parallelize independent sub-tasks (e.g., research 5 libraries at once).
- When noticing that the main agent gets distracted or "hallucinates" due to the noise of accumulated context.
- To implement adversarial review and critique processes (Debate Pattern).
🧠 Architecture Patterns
1. Supervisor / Orchestrator
A central agent controls the flow, decomposes the task, and delegates to specialists.
- Use: Tasks with clear sequential steps and a need for human supervision.
- Risk: The supervisor becomes a context bottleneck.
2. Peer-to-Peer / Swarm
Agents pass control directly to each other through "handoff" protocols.
- Use: Flexible exploration, customer support, dynamic workflows.
- Advantage: Eliminates the supervisor's paraphrasing bias.
3. Hierarchical
Layer structure: Strategic -> Planning -> Execution.
- Use: Massive projects, large-scale code migrations, monorepo management.
🛠️ Implementation in Gemini Elite Core
Pattern: The Direct Pass-Through (Fixing the Telephone Game)
Prevents the supervisor from summarizing the worker's output. Use a tool forward_to_user so that the specialist speaks directly.
<TOOL_CALL: forward_to_user>
{
"message": "[Contenido técnico crudo y preciso generado por el especialista]",
"reason": "Evitar pérdida de fidelidad en la síntesis del supervisor."
}
</TOOL_CALL>
Pattern: Parallel Research Swarm
Launch N agents in parallel and consolidate their findings into a "findings" file.
- Planner: Creates N research tasks.
- Workers: Execute in parallel on isolated sandboxes.
- Consolidator: Reads all results and generates the final report.
🔗 Integrations
context-optimization: Multi-agent partitioning is the purest form of optimization.memory-systems: Use of shared memories (Durable Objects, Disk) for coordination.evaluation: Validation of synthesis quality between agents.
📚 Internal References
references/frameworks.md: Technical implementation details for LangGraph, AutoGen, and CrewAI. Consensus and failure recovery patterns.