What Is Agentic AI? + Prompt Engineering Fundamentals for Agents

Tokens, limits, cost, and time

We will introduce tokens as the units into which text is divided for processing. Tokens give us a practical way to reason about the size of the input, the size of the output, and the maximum context that the model can receive.

Because the prompt is the whole conversation, conversations grow. Every time we submit the conversation again, the model may need to process the information that has accumulated. Longer conversations consume more context, cost more, and generally take longer to process.

So managing a conversation is not housekeeping around the edges of an agent. It is central to agent design. We must continually ask: what information does the LLM need for the next response, and what information is merely occupying space?

Instructions are useful guardrails, but not guarantees

System instructions can establish rules and constraints, but language is semantically flexible. The same underlying request can be expressed in many different ways. A framing that appears obviously problematic in one form may appear benign in another.

We will use a safe classroom example to show how creative reframing changes the model's interpretation. The lesson is not that instructions are useless. It is that instruction-based constraints should not be mistaken for mathematical guarantees. The same flexibility that makes language useful for expressing novel problems makes it difficult to anticipate every representation of an unwanted task.

Multimodal conversations

The conversation does not have to contain only text. A user message can include an image, text plus images, or another supported modality. The image becomes part of the context available to the model.

We will demonstrate prompts that require reasoning about imagery and then return to the same foundational language: the model is not gaining a separate magical sense. We are placing another kind of information into the conversation so that it can influence the next response.

Return observations through images

Finally, we will repeat the cooking loop with photographs. The LLM requests an action, the human performs it, and the human returns an image of the result. The model interprets the image and decides what to request next.

The conversation now contains multimodal observations, but the structure has not changed. One participant requests action; the other participant acts and responds; the new response becomes context for the next turn.