Module 3: Agentic Architecture and Orchestration

Lesson 4: Task Decomposition and Delegation

In the world of AI architecture, "Bigger is rarely Better." If you give an agent a massive, multi-part goal, you are inviting Hallucination. The probability of an error increases exponentially with the number of variables an agent must track simultaneously.

The solution is Decomposition: the process of breaking a "Coarse" goal into "Fine" tasks.

In this lesson, we will learn the heuristics for "Slicing" a project so that an agent can solve it with 100% accuracy.

---

1. What is an "Atomic Task"?

An atomic task is a task that can be completed in a Single Turn with a Single Tool Call and a Clearly Defined Input/Output.

Example:

• Non-Atomic: "Research our competitors and build a dashboard." (Too many variables).
• Atomic 1: "List the top 3 competitors in the European market."
• Atomic 2: "Fetch the annual revenue for Competitor X."
• Atomic 3: "Create a bar chart JSON object using the collected revenue data."

By delegating these "Atomic" steps, you ensure that the model is never "Overloaded."

---

2. The Heuristics of Decomposition

How do you know when to split a task? Use the 3-S Framework:

1. Scope: Does the task require more than one context domain? (e.g., Code AND Marketing). If yes, Split.
2. Steps: Does the task require more than two tool calls? If yes, Split.
3. Success: Can you write a simple Python function to "Test" if this step worked? If the test logic is too complex, the task is too big. Split.

---

3. Delegation Strategies: Context vs. Responsibility

When you delegate a task to a Worker (from Module 3, Lesson 3), you must decide what context to send with it.

• Full Delegation: "Here is the entire history, do Step X." (High risk, high context noise).
• Sub-Atomic Delegation: "Here is ONLY the data needed for Step X. Return ONLY the answer." (Low risk, high focus).

Architect's Tip: Always aim for Sub-Atomic Delegation. By shielding the worker from the "History," you prevent it from being distracted by the Planner's previous mistakes.

---

4. Visualizing Decomposition

graph TD
    Goal[Complex Objective]
    Goal --> T1[Research Task]
    Goal --> T2[Creation Task]
    T1 --> ST1[Search Google]
    T1 --> ST2[Extract PDF]
    T2 --> ST3[Draft Text]
    T2 --> ST4[Format for Web]

By the time the logic reaches ST1 (Sub-task 1), the agent's work is incredibly simple. Simple tasks = Reliable systems.

---

5. Architectural Trap: "The Coordination Tax"

While decomposition increases accuracy, it also increases Latency and Token Overhead. Every time you split a task, you add a "Managerial Turn."

• The Balance: Only decompose until you reach the Reliability threshold. If a single prompt can solve a task with 99% accuracy, do not split it into 5 steps.

---

6. Summary

• Decompose coarse goals into atomic tasks.
• Use the Scope-Steps-Success check to determine split points.
• Sub-Atomic Delegation is the "Privacy Filter" for AI reliability—send only what is needed.

In the next lesson, we will look at how these distributed agents talk to each other: Coordination Strategies Across Agents.

---

Interactive Quiz

1. Define an "Atomic Task" in your own words.
2. Explain the "Coordination Tax" of excessive decomposition.
3. Why does "Shielding" a worker from history increase its reliability?
4. Take a complex goal: "Organize a wedding." List 3 sub-tasks and 6 sub-atomic tasks for this goal.

---

Reference Video: