Agentic workflow or automation: who decided the path?

Take a small support folder. It holds a missed appointment, a question about a quote, and a receipt with no job number.

You can handle that folder in 2 ways. You can write the route before anyone opens it. Or you can let an AI read each item and select a review route inside a boundary you set.

The first route is fixed automation. The second is an agentic workflow.

Neither name tells you whether the work is good. Look at the source file, the rule, the result, and the person who checks the result.

I built a practice folder with 3 fictional requests and 3 written rules. The requests covered a missed appointment, an unclear quote detail, and a receipt without a job number.

The fixed script used those written rules. It matched the appointment to owner review before any fee or message. It marked the quote for owner review because the estimate did not mention furniture. It drafted a request for the missing job number, then marked that item for owner review.

The script did not decide on a new route. It wrote a local Markdown review note.

I then gave the same fictional requests to a Terra coding agent. Terra selected appointment policy for the missed appointment, estimate scope for the quote, and missing receipt detail for the receipt. It wrote its own local review note, terra-output.md.

Both outputs stayed inside the practice folder. Neither result contacted a person, sent a message, changed a booking, charged money, used an account, or wrote outside that folder.

Terra is the name of the AI coding agent I used. Codex is the app that ran Terra on this computer.

The saved note shows one simple result. Terra read the 3 fictional requests, chose 3 review checklists, and wrote a local note. The note did not send a message, change a booking, charge money, use an account, or write outside the practice folder.

The transcript also shows that Terra opened one instruction file from this computer alongside the practice requests. That means this test cannot tell you that the practice folder was the only place Terra could read. It tells you what the saved note did, not every file the agent could reach.

The run also leaves cost, speed, real-world readiness, and wider computer access unanswered. It shows one local result where an AI read different requests and selected from a small list of review checklists.

Write down the inputs you expect. Then write down the route each input should take.

If you can name the inputs and the route before the job starts, fixed automation usually gives you the clearer system. You can test each written rule. The written rule names why a record took its route.

If a new item regularly changes the reading job, an agentic workflow may help. The AI can read the item, choose the matching review checklist, and prepare a local note.

The agentic workflow does not need authority to finish the job. A person can read the source file beside the note and choose the action that reaches someone else.

An AI agent is the worker that reads, chooses, and uses the tools you allow. Agentic AI describes the wider way of working, where that worker chooses steps toward a stated result.

In a small support folder, the difference rarely changes your first decision. You still need a known input, a small allowed action, a reviewable output, and a person for outward actions.

The phrase agentic workflow helps when you need to talk about the whole path. The phrase AI agent helps when you need to talk about the worker inside the path.

The first result should be a note you can check beside the source files. Put the original request on one side and the route the system chose on the other.

For the missed appointment, check the policy before you accept the review note. For the quote question, check the estimate before you accept a scope decision. For the receipt, check that the system asked for the missing job number instead of pretending it knew one.

This habit works for both approaches. A fixed route can carry a bad rule. An agentic route can choose a poor checklist. The source files show you what happened.

Choose a copied pile with 3 to 10 items. Remove customer details. Write the fixed route on one page. Write the review checklists on another page.

Run the fixed route first. Then run an agentic route only inside that copied folder. Compare each local note with the original item.

Claude Code did not produce the earlier Terra comparison. The free Claude Code 101 course is Anthropic's official introduction to Claude Code. The practice in this post is separate: use a small local folder, inspect the result, and keep a person in charge of outside actions. Hands-on Claude Code needs Pro, Max, Enterprise, or API access, so check your access route before planning a practice run.

When you want a broader guided route after that first folder, take my AI Agents course.