AI Agent

An AI agent is a system in which a language model does not just answer once but pursues a goal over multiple steps — deciding what to do, calling tools (search, code execution, file edits, APIs), observing results, and iterating until the task is complete or it asks for help.

The distinction from a chatbot is the loop. A chatbot maps one prompt to one response. An agent runs a cycle — reason about the goal, pick an action, execute it, read the outcome, decide again — with the LLM as the decision-maker and tools as its hands.

How it works

Four components recur in almost every agent design:

• The model: the reasoning core that plans and chooses actions.
• Tools: functions the model may invoke, each described by a schema — run a query, call an API, edit a file, browse a page. The model emits a structured tool call; the surrounding harness executes it and feeds the result back.
• Memory: the running context of the task — what has been tried, what was learned — sometimes extended with persistent storage or retrieval across sessions.
• The loop and its guardrails: step limits, budget caps, permission gates on dangerous actions, and logging of every decision.

Agent autonomy is a dial, not a switch. At the cautious end sits the propose-and-approve pattern: the agent plans and drafts actions, but a human approves each consequential step before it executes. At the far end, fully autonomous agents act unsupervised — appropriate only where mistakes are cheap and reversible.

Why it matters

Agents move AI from answering questions to doing work: triaging support tickets end to end, researching across dozens of sources, writing and testing code, operating a browser. But the loop also compounds risk — a hallucination in step 2 becomes a wrong action in step 3, and an agent with production credentials can do real damage at machine speed. The engineering discipline around agents — tight tool permissions, audit logs, human gates, and evals of whole task trajectories rather than single responses — matters more than the model inside them.

A worked example

Goal: "Find why the nightly data sync failed and propose a fix."

Step 1: Agent calls the log-search tool for last night's sync job;

finds repeated timeout errors at 02:13.

Step 2: Calls the metrics tool; sees database CPU pegged at 100

percent in the same window.

Step 3: Greps the repo; finds a query added last week missing an

index on the join column.

Step 4: Drafts a migration adding the index, plus a summary linking

log lines to the diff — and stops, awaiting human approval.

Reasoning, tools, evidence, a proposal — and a human gate before anything touches production.

How miatz teaches and uses it

miatz learners build agents from parts — first a bare tool-calling loop, then memory, then guardrails — and War Room drills include agent failure incidents drawn from documented outages. The institute also runs on one: Xook, the AI shadow engineer, operates strictly propose-and-approve with every action audit-logged. Learners watch a production agent work safely every day, which teaches more than any lecture on agent safety could.