Architectural Trade-offs in Agentic Systems
The research investigates how to effectively structure AI agents for long-horizon tasks—complex objectives that require multi-step reasoning and execution over extended periods. The core tension lies between two primary paradigms: Subagents and Agent Skills.
The Subagent Pattern: Hierarchical Delegation
Subagents function as independent, specialized agents invoked by a primary controller. This approach mimics a hierarchical management structure where the main agent delegates entire sub-tasks to a subagent that maintains its own context, prompt, and toolset.
- Advantages: High modularity and isolation. Each subagent can be optimized for a specific domain, reducing the cognitive load on the primary model.
- Trade-offs: Significant overhead in context management and communication latency. Passing state between the controller and subagents can lead to information loss or "context drift" if not strictly managed.
The Agent Skill Pattern: Reusable Procedural Knowledge
In contrast, the Agent Skill pattern treats complex operations as reusable, atomic functions or "skills" that the primary agent can call directly. This shifts the complexity from architectural hierarchy to the agent's internal reasoning capabilities.
- Advantages: Lower latency and reduced context switching. Because skills are integrated into the primary agent's execution loop, the agent maintains a more coherent global state.
- Trade-offs: Requires a highly capable primary model that can effectively manage a large library of skills without becoming overwhelmed by the choice of which tool to invoke (the "tool selection" bottleneck).
Strategic Implementation for Long-Horizon Tasks
The findings suggest that the choice between these two patterns should be dictated by the nature of the task:
- Task Complexity vs. Frequency: If a task is highly complex but rarely performed, the subagent pattern provides better isolation and debugging. If a task is a frequent, repeatable operation, encapsulating it as a skill is more efficient.
- State Persistence: Long-horizon tasks often fail due to state corruption. The research highlights that skill-based architectures generally perform better when the task requires constant, real-time access to the global state, whereas subagents are superior for "fire-and-forget" sub-tasks that operate within a bounded context.
Ultimately, the authors argue that modern agentic systems should likely adopt a hybrid approach: using subagents for high-level planning and orchestration, while utilizing a library of optimized skills for the granular execution of repeated technical steps.