Generative AI creates content, such as text, images or code, from a prompt. Agentic AI goes further: an AI agent uses a model to plan steps, call tools and APIs, check results and keep going until it completes a goal. For developers, agents are software systems built around an LLM, not magic.
What is generative AI?
Generative AI models produce new content based on patterns learned from data. Large language models write and summarise text and code; image models create pictures. A chatbot answering a question is generative AI.
What is agentic AI?
An AI agent wraps a model in a loop: read the goal, decide the next step, use a tool (search, a database query, an API call, running code), look at the result, and repeat. The model provides the reasoning; your code provides the tools, limits and memory.
| Generative AI | Agentic AI | |
|---|---|---|
| Output | Content (text, code, images) | Completed tasks and actions |
| Steps | Usually one response | Many steps in a loop |
| Uses tools | Optional | Central |
| Main risk | Wrong or made-up content | Wrong actions with real effects |
The building blocks of an AI agent
- A capable model that can follow instructions and choose between options.
- Tools: well-described functions the agent may call, each with narrow permissions.
- Retrieval and memory: access to relevant documents (RAG) and a record of what has happened.
- Planning and control: a loop with step limits, timeouts and clear stopping rules.
- Evaluation: test cases that check whether the agent actually completes tasks correctly.
- Guardrails: input checks, output checks, human approval for risky actions, and defences against prompt injection.
Where are agents useful today?
Agents work best on well-defined, repetitive tasks with clear success criteria: triaging support tickets, pulling data from several systems into a report, running routine code changes with tests, or filling structured forms from documents. The more open-ended and high-stakes the task, the more human review you need.
How to learn to build agents
Start with strong software fundamentals, because an agent is mostly careful software. Then learn how LLMs work (explained here), prompting, tool calling, retrieval and evaluation. Program Zero covers agents and tool use in Phase 8, and prompt injection and guardrails in the security phase.
