What an AI agent actually is, for people tired of the hype

"Agentic AI" is doing a lot of work in job descriptions right now, and very little of that work is definitional. Here is the version that survives contact with a real codebase.

An agent is a loop with tools

That is genuinely most of it:

1. A model receives a goal and the current state.
2. It picks a tool to call — search a database, send an email, run a query.
3. The tool runs and returns a result.
4. The result goes back into the model, and the loop repeats until the goal is met.

A chatbot answers. An agent acts, checks what happened, and acts again.

The four parts you actually build

Tools. Ordinary functions with a described input schema. get_invoice(invoice_id). The hard part is not writing them, it is describing them precisely enough that the model calls them correctly.

Memory/context. What the model can see. Too little and it flails, too much and it gets expensive and slow. This is where retrieval comes in.

Guardrails. What the agent is not allowed to do. An agent with database write access and no guardrails is an outage waiting to happen.

Evaluation. How you know it works. A test suite of realistic tasks with expected outcomes. Teams that skip this ship agents that regress silently.

What people are being hired to build

Not chatbots. The requests we see from companies hiring our students:

• Invoice and document processing that reads PDFs and files them into an ERP
• Customer support triage that classifies and routes tickets, escalating what it cannot handle
• Internal search over a company's own documents, with citations
• Data-entry replacement that reads emails and updates a CRM

Every one of these is boring, valuable, and mostly plumbing.

What to learn, in order

1. Python well enough to build an API. Not "hello world" Python.
2. Calling a model with structured output and tools.
3. Retrieval: embeddings, chunking, why naive chunking fails.
4. Evaluation: write the test set before you tune the prompt.
5. Deployment, cost and latency.

Notice that only two of the five are about the model. The rest is engineering, which is why engineers are the ones getting these jobs.

The honest caveat

Agents fail in ways ordinary software does not — plausibly, and without an error message. Anybody selling you a course that does not spend real time on evaluation and guardrails is selling you the demo, not the job.