Is It Actually an AI Agent? The Five-Point Test

Quick Answer: What Counts as an AI Agent?

Software is an AI agent if it decides which steps a goal requires, runs those steps itself through tools that read real data and produce real artefacts, and can reach systems outside the chat window. If you still choose every step and paste the output somewhere, it is an AI feature with an agent's name on it.

That definition is doing real work in 2026, because the word has been applied to almost everything. This post turns it into five checks you can run on a vendor before a trial ends.

Why "agent" stopped meaning anything

Read the current roundups and the same complaint opens most of them: every social tool released this year calls itself an agent, whether it is a scheduler, a chatbot builder, a caption generator or an analytics dashboard. The tested comparisons say it plainly — most tools marketed as agents generate a caption when you press a button, suggest a posting time, and stop there.

The industry has a name for this now: agent washing. It matters commercially, because the two products are priced alike and behave nothing alike. A caption generator saves you the blank page. An agent changes who does the assembly.

What follows is deliberately mechanical. You do not need a definition everyone agrees on; you need five questions with answers you can check.

The five-point test

#	Check	The question to ask	Fails if
1	Tool count	How many tools can it call, and how many of them read rather than write?	There is one capability — generate text — with a chat window in front of it
2	Chain depth	How many steps will it run from one instruction before it needs you again?	One instruction produces exactly one action
3	Grounding	Which of your data does it read before it writes anything?	It reads your prompt and nothing else
4	Action authority	Can it change something outside the conversation — a queue, a calendar, a live post?	It hands back text and you carry it somewhere
5	Approval model	Is the stop before publishing a code path, or a sentence in a prompt?	Nobody at the vendor can answer this question

1. Tool count

An agent needs things to do other than write. The useful split is between reading tools (analytics, connected accounts, the posting queue, a product catalog, the live web) and producing tools (copy, images, video, edits). A product with ten generation tools and no reading tools is a content factory, which is a legitimate thing to buy — just not an agent.

Ask for the list. If the documentation has no such list, that is your answer.

2. Chain depth

Depth is how many steps the software will take on its own before handing back. One is an AI feature. Three to ten is an agent. The number itself matters less than whether the vendor can state it, because a vendor who has implemented tool chaining knows their limit and a vendor who has not will change the subject.

3. Grounding

This is the check that catches the most convincing impostors. Ask the tool for something it cannot answer without looking: "Take last month's best performing post and draft three more like it for the product with the highest price."

A generator will answer confidently and invent both the post and the product. An agent will fetch, and will tell you what it fetched. Run this once and the category resolves itself.

4. Action authority

Can it put something in the queue? On the calendar? In public? This is the line where a helpful writing tool becomes a system with consequences, and it is also where the price difference between the two categories is justified.

5. Approval model — the check that outranks the other four

Four out of five is not the passing grade. The fifth is, because it is the only one that describes what happens when the software is wrong.

There are three answers a vendor can give, and they are not equivalent:

• "It publishes autonomously." Someone has decided that a model's judgement is sufficient for a public brand action. You now own that bet.
• "We instruct it to ask first." A prompt-level rule is a request. A model under a long chain of instructions can be argued out of a request, and prompt injection exists precisely to do that.
• "The publish path requires an approval that cannot be disabled." A gate in the execution path is not a request. It is the only one of the three that holds when the model is wrong.

Ask which one it is. Then try to turn it off in the settings, because the answer to that is more honest than the answer to the question.

How to run the test in ten minutes

1. Count the tools. Find the tool or capability list in the documentation and count how many read your data rather than generate text. A generator has one capability; an agent has a registry.
2. Send a request that needs two steps. Ask for something it cannot answer without fetching something first — last month's best post, the newest product, the queue for next week.
3. Check whether it fetched or invented. Ask where the numbers came from. An agent cites what it read; a generator produces plausible numbers that are not real.
4. Ask where the approval boundary is enforced. Code path, or system prompt? Put it in writing to support if you have to.
5. Try to switch the approval off. If a setting removes it, the safety of the account is now an operational problem you own.

Scoring

Score	What you are buying
4–5	A tool-using agent. Evaluate it on the boundary, not the feature list.
2–3	An assistant with a couple of integrations. Often the right purchase — just do not pay agent prices for it.
0–1	A generator with a chat window. Useful, cheap, and mislabelled.

The three failure modes the test catches

The generator in a trench coat. Scores 0–1. One capability, no reading, no actions. Catches on check 3 every time.

The workflow builder wearing the word. Scores 2–4 but fails the spirit of check 2: the steps are real, but you defined every one of them in a canvas. That is a workflow, and a workflow is often the better tool — it is repeatable and it never improvises. It is simply not an agent. The distinction is worth its own read: AI agent vs AI workflow.

The agent with no brakes. Scores 4 on the first four checks and fails the fifth. This is the dangerous one, because it demos better than anything else in the category.

How Autoadify scores on its own test

Publishing a rubric obliges us to run it on ourselves, including the parts that do not flatter.

• Tool count — pass. Twenty tools. Eleven read (connected accounts, past posts, the scheduled queue, engagement analytics, the Shopify or WooCommerce catalog, stock libraries, Canva designs, the live web), six generate (copy, images, video, edits), three act.
• Chain depth — pass. Up to ten steps in a single request.
• Grounding — pass. It reads the catalog and the analytics before it writes, which is the entire reason the output is about your products rather than about products in general.
• Action authority — pass. Three action tools: schedule a post, schedule a whole content plan, publish now.
• Approval model — pass, by the strict definition. Those three tools, and only those, stop for explicit human approval. It is enforced in the execution path and no setting removes it.

And the honest limits, which the test does not measure: the agent is not autonomous. It works on request, in the workspace — it does not sit in the background overnight watching your accounts. It does not answer comments or DMs. Publishing that genuinely runs unattended is a different feature with a different name, AI Workflows, which fire on a trigger you configured in advance.

Any vendor answering all five checks should be equally specific about what their agent does not do. Vagueness there is its own signal.

Frequently Asked Questions

What is agent washing?

Marketing an existing AI feature — usually a caption or image generator — as an autonomous agent. It is the dominant pattern in social media tooling in 2026, and it is why a feature checklist no longer distinguishes products in this category.

Is a chatbot an AI agent?

Not on its own. A chatbot converses. An agent calls tools, reads real data and changes something outside the conversation. A chatbot with tools attached can be an agent; a chatbot alone is check 1 failed.

Does an AI agent have to be autonomous?

No, and conflating the two is the most expensive mistake in this category. Autonomy describes whether a human is in the loop. Agency describes whether the software can plan and act at all. A tool that plans ten steps and then stops for your approval is fully an agent and deliberately not autonomous.

How many tools does an AI agent need?

There is no threshold, but the ratio matters: an agent that can only generate is a generator. Look for reading tools, because reading is what makes the output specific to you.

Can I run this test during a free trial?

Checks 2, 3 and 5 need nothing but a trial account and one deliberately awkward request. Checks 1 and 4 are answered by the documentation.

Which check matters most for a brand account?

The fifth. The first four describe what the software can do for you; the fifth describes what it can do to you.

See it work

Autoadify's agent carries twenty tools, chains them up to ten deep, and stops for approval on every action that reaches the public. See how the agent works, or start free and run the five checks on it yourself.