Assistant or agent? What AI can actually do for you today

Picture Tuesday evening as you plan dinner for four people: vegetable lasagna, budget up to 35 euros.

• An AI assistant answers your question directly in the chat: it suggests a recipe and writes out a tidy shopping list. Then it waits for you.
• An AI agent is given a goal and connected tools: it checks your pantry note, searches the online grocery store for missing ingredients, sees that organic ricotta is out of stock, picks a suitable alternative within budget and places the items into the cart, pausing before payment for your approval.

The technical difference is not a magic buzzword, but the work loop with tools:

1. Understand the goal and plan a step: The model breaks the assignment into checkable sub-steps.
2. Call a tool: Through a defined interface (such as function calling or the Model Context Protocol, MCP), the system sends a structured request to a program, such as a search service, calendar or shopping cart.
3. Inspect feedback: The tool returns a real result or error message ("Ricotta unavailable").
4. Adapt the next step: The agent adjusts its plan based on that feedback until the goal is reached or an approval boundary stops it.

• Anthropic Engineering: Building Effective Agents
• Model Context Protocol (MCP) Architecture

Greater autonomy brings two typical risks:

• Error chains: If an assistant misunderstands an ingredient, you spot it right away in the reply. If an agent makes a flawed assumption in step 2 and builds three more tool calls on top of it, the wrong item can end up in your cart or the wrong meeting on your calendar.

• Indirect prompt injection: When an agent reads external websites, emails or documents while researching, manipulated text inside those sources can try to slip unauthorized commands into the agent's loop.

• OpenAI: Function Calling Guide

• OWASP Top 10 for Large Language Model Applications (LLM01: Prompt Injection)

Three simple rules help decide when a plain assistant is enough and how to bound an agent safely:

1. Allow reading and drafting freely: Searching, summarizing, comparing and preparing drafts can be automated safely because nothing is changed in the outside world yet.
2. Put a stop sign before writing, booking or sending: Wherever money moves, messages go out to third parties or data is deleted, let the agent prepare the step only up to a preview. You approve the final click yourself.
3. Connect only the tools needed for the task: A grocery agent needs access to recipes and the cart, not to your email inbox or bank transfers.