Automations vs. agents: which one do you actually need?
Two different machines. An automation (also called a workflow: a fixed set of steps that runs the same way every time) is a train on rails. An agent (an AI that decides which steps to take as it goes) is a taxi driver. The train is fast, cheap, and predictable, but it can only go where the track goes. The driver can handle a messy, never-seen-before route, but costs more and makes judgment calls.
🚂 Automation / workflow
Same fixed steps, every single time.
Trigger→Step A→Step B→Done
🧭 Agent
Looks at the input, then decides what to do.
Goal→Decide→Tool?→Re-check
🚂 Automation: predictable & cheap🧭 Agent: flexible & adaptive🛤️ Best fit: does the path change?🤝 Often both at once
Keep one question in your pocket the whole way through: does the path change from job to job? If it never changes, you want rails. If every job is a little different and needs judgment, you want a driver. In a minute, you will sort real tasks yourself.
02 / 05 · The real difference
The myth: an agent is always the upgrade.
This is the costly mistake. People hear "agent" and assume it is the smarter, better, more advanced version of an automation. It is not a tier above. It is a different tool for a different job.
⚠ The myth
"Automations are old. If we are doing it properly, we should use an agent."
✓ The truth
An agent is not better, it is looser. You hand control to a model that decides what to do. That is exactly what you want when the input is messy and the path is unknown, and exactly what you do not want when the steps are fixed. For a job that never changes, an agent is slower, more expensive, and less predictable than a plain automation that does the same five steps every time. You added judgment to a task that needed none.
🚂 Reach for an automation when
✓ The steps are the same every time
✓ The input is clean and structured (a form, a row, a file)
✓ You need it cheap, fast, and identical on every run
✓ A wrong guess would be costly, so you want zero guessing
🧭 Reach for an agent when
✓ The input is messy or free-form (an email, a photo, a chat)
✓ The right next step depends on what was found
✓ The task needs judgment, not just a rule
✓ You cannot list every case in advance
The key: automations are cheaper and safer precisely because they cannot think. Agents are useful precisely because they can. Picking the wrong one does not just waste money, it makes a reliable task flaky, or a flexible task rigid.
03 / 05 · Which fits?
Pick a task. Watch it get sorted.
Tap a real task below, or type your own. The page reads it and calls it an automation, an agent, or a hybrid, and tells you exactly why. This runs entirely in your browser; nothing is sent anywhere.
🔒 100% client-side. No network call, no AI, no API key. Just a few keyword rules running in your browser, so you can see the reasoning, not a black box.
💡 The rules here are deliberately simple. A real Rabbithole build looks at your actual data and volume before choosing. The point is the mental model, not this toy classifier.
04 / 05 · The hybrid
Most good systems are both.
The real answer is rarely "pure automation" or "pure agent." The strongest builds put an agent only on the fuzzy part, and keep deterministic rails around everything else. Deterministic just means: given the same input, you always get the same output, with no guessing.
Example: an inbox that handles support email.
Rails (automation): an email arrives, gets logged, and is filed. Same every time, no judgment needed.
Judgment (agent): read this specific, messy message and decide what it is actually asking for. This is the fuzzy part, so it gets a model.
Rails again (automation): route to the right queue, attach the order record, draft a reply for a human to approve. Fixed steps once the intent is known.
The agent handles the one step a rule cannot: understanding plain-English intent. Everything around it stays on rails, which keeps the whole thing cheap, fast, and auditable.
A safety note: the more you let an agent decide, the more it can do, and the more it can get wrong. Put rails on anything that costs money, sends a message, or cannot be undone. A common pattern is to let the agent decide and draft, but require a human to approve before anything real happens. (For more on this, the AI safety lesson covers it.)
05 / 05 · Done
You now understand this better than most people who build automations daily.
You know that an automation runs fixed steps and an agent decides its own, that agents are not an upgrade but a different tool, and that the best systems are usually a hybrid: reliable rails with a model on the fuzzy part only.