Not every workflow that uses AI agents is truly agent-driven.

You can build a workflow that calls multiple AI agents, yet still control every step through code. You can also give an agent a goal, a set of tools, and the freedom to determine how to complete the task.

Both approaches are useful, but they solve different problems.

In this session, we'll explore the difference between workflows that use agents and workflows where agents drive execution.

We'll build the same incident investigation system using two architectures: a predefined workflow with specialized AI agents and an agent-driven workflow that dynamically selects tools, investigates problems, and adapts its plan.

Then we'll introduce failures, inspect execution traces, and compare the results across reliability, cost, latency, and control.

We'll also explore hybrid architectures that combine deterministic workflows with bounded agent autonomy, human approval, and evaluation.

You'll learn to:

1. Distinguish code-driven orchestration from agent-directed execution.

2. Implement and compare workflow, agent-driven, and hybrid architectures.

3. Choose the right level of autonomy based on predictability, reliability, cost, and operational risk.

The question isn't how many agents your system needs. It's how much control you should give them.