When to use an agent, and when not to

Agents are a tool, not a goal

Agents are powerful and also slower, costlier and less predictable than ordinary software. Good builders pick the simplest thing that works.

A decision guide

Use plain automation or a script when the steps are known and never change, the inputs are structured, and a mistake is costly. Paying an invoice from an approved list is a good example.

Use a single model call (a chatbot-style task) when you only need a transformation of text you already have: summarise this, translate that, classify this message.

Use an agent when the job needs judgement over several steps and the steps depend on what is found along the way. Examples: "read this week's support tickets, look up the account for each unusual one and tell me which need a human", or "check these three sources and write a briefing that mentions only what changed."

Questions to ask first

1. Is the path fixed? If yes, automation is cheaper and more reliable.
2. What is the cost of a wrong answer? If a mistake could harm someone, make a person approve the final action.
3. Can I tell if it worked? If you cannot check the result, you cannot trust it. Write down what a good output looks like.
4. Is the job narrow enough? One clear job per agent is easier to test and to trust than a do-everything assistant.
5. How often will it run? A job that runs thousands of times needs a close look at cost per run.

Signs an agent is the wrong choice

• You need the same answer every time, exactly.
• The task involves money movement, legal commitments or health decisions with nobody reviewing.
• The data it would need cannot be shared with it.
• You cannot describe success.

How this looks in PGB Agents

PGB Agents supports the safe middle ground: agents that gather and draft on their own, but must ask before they send, share, spend or install. Many good first agents are read-and-report jobs with no write tools at all.

Try it yourself

1. Write down three tasks from your own work or life.
2. For each, choose: automation, a single model call, or an agent, and note the one-sentence reason.
3. Open New agent and scroll to From the marketplace. Pick one listed agent and read its summary. Which of the five questions does it answer well?

Recap

• Choose the simplest approach: automation for fixed paths, a model call for text transforms, an agent for multi-step judgement.
• Ask about path, cost of error, checkability, scope and frequency.
• Keep a person in the loop wherever a mistake would be expensive.