Forward-deployed / Learning zone
Agentic AIa standalone module
Lesson 01

What is an agent?

TL;DR

An agent is a language model given tools, a goal, and permission to loop. It gathers context, decides an action, executes it, observes the result, and repeats until it meets the goal or gives up. That loop is the essence — not any particular framework, protocol, or "layer." Everything else in agentic AI attaches to the loop: memory feeds it, planning steers it, evals grade it, sandboxes contain it. The key product variable is autonomy: a spectrum from a fixed workflow that calls a model at predefined steps, to an agent that decides its own steps. More autonomy buys flexibility on open-ended problems. It also costs you predictability, latency, money, and debuggability. That's why the first agent design decision is always how little autonomy can we get away with?

🎯 For the AI PM

Why it matters — "Agent" is the most abused word in the industry. People apply it to everything from a cron job with an LLM call to a fully autonomous coding system. If you can't place a proposal on the autonomy spectrum, you can't estimate its cost, risk, or failure modes. And you can't compare two "agents" that differ by an order of magnitude in both.

What it changes in your decisions — For every "let's build an agent" proposal, ask first whether a fixed workflow with LLM steps would do the job. A workflow is cheaper, faster, more predictable, and easier to debug. Reserve full agency for problems where you genuinely can't enumerate the steps in advance.

Ask yourself — "Could I draw this task as a flowchart? If yes, why am I paying for an agent to rediscover the flowchart on every request?"

Risk if ignored — You ship an expensive, unpredictable loop where a five-step pipeline would have done the job. Or you demo a scripted workflow as an "agent" and set expectations autonomy can't meet.

The loop

Strip away every framework and an agent is this:

The agent loop

Gather, decide, act, observe, check · everything else attaches to this loop

Strip away every framework and an agent is this cycle, with budgets and exits.

Goal — user intent
Gather context — instructions, history, retrieved knowledge
Model decides next action
Act — search, run code, call an API, write
Observe result — append to context
Goal met? Budget left?
No → loopback to "model decides"
Stuck / risky → escalatehand off to a human
Yes → respond / deliverthe loop ends

On each pass through the loop, the model sees everything so far: the goal, its previous actions, and their results. Then it picks the next move. Three properties follow directly from this shape, and they explain most of agent behavior:

The canonical anatomy: model, tools, orchestration

Google's 2024 Agents whitepaper is the closest thing the industry has to a shared reference. It names the three components inside every serious implementation of the loop. The vocabulary is worth adopting, because your engineering team already uses it:

The same paper draws a useful model-vs-agent contrast. A model's knowledge ends at its training data; an agent extends it live through tools. A model does one inference per query with no session state; an agent manages multi-turn history in the orchestration layer. A model has no native tool use or logic layer; an agent has both built in. If a proposal calls something an "agent" but has no tools, no session management, and no loop, it's a model call wearing a costume.

The autonomy spectrum

"Agent or not" is the wrong question. How much agency is the right one:

Level What decides the steps Example Failure style
LLM call Nothing — one shot Summarize this ticket Wrong output
Workflow Your code; LLM fills slots Extract → classify → route pipeline Predictable, per-step
Router LLM picks a branch, code runs it Support triage choosing a playbook Wrong branch
Tool-using agent LLM picks actions from a toolbox, loops Research assistant with search + docs Wanders, over-calls tools
Autonomous agent LLM plans, acts, self-corrects over long horizons Coding agent fixing a bug end-to-end Expensive, novel failures

Two heuristics carry most of the decision. Enumerable steps → workflow. If an expert can write down the procedure, encode the procedure. Use the model inside the steps only where judgment or language handling is needed. Workflows are cheaper, faster, testable, and their failures stay local. Open-ended path → agent. Sometimes the steps genuinely depend on what you discover along the way — debugging, research, negotiation. A fixed pipeline then either explodes into unmanageable branches or fails on the first surprise. That's what the loop is for.

A third heuristic comes from hard experience: match autonomy to stakes and reversibility. An agent drafting replies a human sends can be boldly autonomous. An agent with delete access to production data had better not be (safety & security).

A loop, not a layer cake

You'll meet diagrams that present agentic AI as a tidy tower: infrastructure at the bottom, then protocols, tools, cognition, memory, applications, and governance stacked on top. They're useful as inventories — most of the named pieces exist. But they're misleading as architecture. The pieces aren't layers with dependencies. The model itself is usually missing from the picture. "Cognition" is mostly the model plus prompting patterns, not an installable component. And governance and observability are cross-cutting disciplines, not a penthouse floor. When you see such a stack, mentally rearrange it into the knowledge graph from the module overview: a loop at the center, capabilities feeding it, disciplines bounding it. That picture predicts behavior. The tower doesn't.

Failure modes

Practitioner checklist