Calling for Backup

Opening Scene

A line cook running the grill station doesn’t also try to personally handle pastry, sauces, and plating — a well-run kitchen delegates each specialized task to the station actually equipped and trained for it, coordinated by someone tracking the whole picture. An AI agent facing a genuinely complex, multi-domain task benefits from this exact same specialization: delegating a specific subtask to a sub-agent purpose-built or specifically prompted for that particular kind of work, rather than trying to be equally capable at everything itself.

In Plain English

Delegation (or sub-agent architecture) has a primary, orchestrating agent break a complex task into subtasks and hand each one to a specialized sub-agent — one focused on research, another on writing, another on code execution — each potentially using a different tool set, a different system prompt, or even a different underlying model chosen for that specific kind of work. This connects directly to the task decomposition covered in Article 5, but extends it: instead of the same agent executing every decomposed step itself, different steps get routed to agents specifically suited to them.

The Old Way

Before delegation patterns matured, agentic systems often relied on one generalist agent trying to handle every kind of subtask itself:

• Early agentic experiments frequently used a single agent configuration to handle every subtask in a plan, regardless of how different those subtasks actually were.
• A general-purpose agent, tasked with something highly specialized, often performed worse than a version specifically prompted or configured for that narrower task.
• There wasn’t yet a well-established pattern for how a primary agent should hand off work to, and receive results back from, a specialized sub-agent.

Delegation architectures emerged specifically as practitioners recognized that specialization, much like in any well-run organization, often outperforms generalism for genuinely varied, complex work.

What’s Changing (and Why AI Is the Reason)

1. Agent frameworks increasingly provide built-in support for orchestrator-and-sub-agent architectures, making this pattern considerably more accessible than custom, from-scratch implementation.
2. This connects directly to the multi-agent systems covered in Article 16, where delegation extends into genuinely distributed collaboration between multiple, more independent agents.
3. As tasks handled by agentic systems have grown more complex and varied, delegation has become an increasingly essential pattern for maintaining both quality and manageable system complexity.

The Metaphor, Fully Extended

The Kitchen	Delegation Concept
A grill cook not personally handling pastry and plating too	A primary agent not personally handling every kind of specialized subtask
Delegating each specialized task to the station equipped for it	Delegating each subtask to a sub-agent configured for it
A kitchen coordinator tracking the whole picture across stations	An orchestrating agent tracking progress across delegated sub-agents
Specialization producing better results than one generalist doing everything	Specialized sub-agents producing better results than one generalist agent

Quick Recap

• Delegation has a primary agent break a task into subtasks and hand each to a specialized sub-agent.
• Specialized sub-agents, configured for a narrower kind of work, often outperform one generalist agent handling everything.
• Agent frameworks increasingly provide built-in support for this orchestrator-and-sub-agent pattern.
• This pattern is an essential stepping stone toward the more distributed multi-agent systems covered later in this series.

Where This Fits in the Series

Article 7 covered delegating specialized work to a sub-agent. Article 8 covers a discipline every agent needs regardless of who does the work: checking it before it goes out.