Transformers: The Definitive Guide

Chapter 9. From Scripts to Thinking: AI Agents for Complex Tasks

Traditional workflows are like well-rehearsed plays: the sequence of actions is fixed, the actors stick to their lines, and the ending is predetermined. Agentic systems break this pattern. Agents, by contrast, are improvisers. They can decide which scene to play next, change their lines on the fly, and even invite new actors onto the stage. In improv, there is the “Yes, and…” rule: accept what is given and add to it so the story moves forward. In AI terms, it becomes: Yes → accept the current state, context, or contribution from another agent or tool without discarding it; and → build on it by adding reasoning, actions, or new information that moves the task forward. Instead of marching in a straight and inflexible line, agentic systems loop through cycles of planning, action, and self-correction. This shift from predefined scripts to adaptive, collaborative decision making lies at the heart of autonomy in AI agents. Table 9-1 compares both workflows.

Table 9-1. Comparison of traditional and agentic workflow

Traditional workflow	Agentic system
Predefined sequence	Dynamic decision loop
Limited flexibility	Context-aware, tool-using
Linear execution	Iterative, self-correcting
No adaptation to feedback	Adapts plans based on feedback
Fixed toolset	Selects and combines tools as needed
Single-pass completion	Can revisit and refine earlier steps
Static goal interpretation	Can reinterpret goals based on ...