The one idea
An agent decides what to do next. A workflow follows steps you already decided.
That is the whole distinction, and it determines everything: cost, speed, predictability, and what happens on a bad day. Flexibility is not free — you pay for it in the ability to know in advance what the system will do.
A normal automation follows your instructions. An agent works out its own instructions as it goes. That is useful sometimes and alarming often.
Use an agent when the steps genuinely cannot be known in advance. If you can write the steps down, write them down — it will be cheaper, faster and it will do the same thing every time.
An agent is a model in a loop with tool access, planning over an open action space toward a goal. Non-determinism is the feature and the risk; every tool granted expands the blast radius of a wrong decision.
What an agent is made of
| Part | What it means | Where it goes wrong |
|---|---|---|
| Goal | What you asked for, in words | Vague goals produce creative interpretations |
| Tools | What it can actually do — send, read, write, run | Every tool is a permission you granted |
| Actions | Its choice of which tool to use, with what input | Right tool, wrong argument, is the common failure |
| Memory | What it carries between steps | Fills up, drifts, remembers a wrong conclusion |
| Planning | Deciding the sequence itself | Plans that loop, or that skip verification |
| Permissions | What it may do without asking | The only real control you have |
Read that table as a risk list. An agent's capability and its danger are the same column: the tools you gave it.
Where an agent genuinely helps
Agents earn their cost in a narrow band, and it is worth naming it precisely.
The steps cannot be known in advance. Researching a question where what you find determines what you look for next. A fixed workflow cannot express that.
The input is unstructured and varied. Reading a hundred differently formatted documents to extract the same three facts. Writing rules for every format is harder than the problem.
Failure is cheap and a human reviews the output. Drafting, summarising, suggesting. If it produces something wrong, a person notices before it matters.
Exploration where being wrong is fine. Generating options, first drafts, possibilities you will filter.
Notice what these have in common: the agent's output is read by a person before it does anything irreversible. That is the pattern.
Where a workflow is better
Fact A deterministic workflow does the same thing every time given the same input. An
agent may not, because the underlying model is probabilistic.
Recommendation If you can write the steps down, use a workflow. Cheaper per run, faster,
testable, auditable, and it fails in ways you can predict.
| Workflow | Agent | |
|---|---|---|
| Same input twice | Same output | Possibly different output |
| Cost per run | Fixed and small | Varies; a loop can be expensive |
| Speed | Milliseconds to seconds | Seconds to minutes |
| Testable | Yes, thoroughly | Only statistically |
| When it fails | Stops at a known step | May carry on, plausibly |
| Explaining what it did | Read the steps | Read the log and hope |
"Move invoices from email to the accounts folder and log the amount" is a workflow. The steps are known, the format is fixed, and being wrong costs money. An agent here adds cost and unpredictability to a solved problem.
"Read these 200 supplier contracts and list the ones with an auto-renewal clause" is agent-shaped. The documents differ, the clause is worded differently every time, and a human reads the resulting list before anyone acts on it.
A team replaced a working rules-based ticket router with an agent. Accuracy went up slightly. Cost went up substantially, latency went from instant to several seconds, and when a ticket was misrouted nobody could explain why.
The rules version was worse at the edges and better at everything else. This is the common shape of the trade, and it is worth measuring rather than assuming.
The failure modes
Five ways agents fail, and what to do
1 of 5Loops.
The agent tries something, it does not work, it tries a variation, and again. It can run for a long time making no progress, and each step costs.
Defence: hard caps. Maximum steps, maximum time, maximum spend. Not a suggestion in the prompt — an actual limit enforced outside the agent, because a limit the agent is asked to respect is a limit it can reason its way past.
Approval gates
The practical control is deciding, in advance, which actions an agent may take alone.
| Risk | Example | Gate |
|---|---|---|
| Reversible, private | Draft a summary, search, read | None needed |
| Reversible, visible | Create a draft, add a tag, open a ticket | Notify afterwards |
| Hard to reverse | Send an email, post publicly, change a record | Human approves first |
| Irreversible or costly | Delete, pay, deploy, message a customer | Human approves, and a second check |
The correct question is never "can the agent do this?" It is "what happens if it does this at the wrong moment, and who finds out?"
Try this
For each, decide: agent, workflow, or neither.
- Turn 500 meeting recordings into summaries for a person to review.
- Approve refunds under ₹2,000 automatically.
- Research five suppliers and produce a comparison for a human decision.
- Rename and file 300 documents by a rule from their filename.
Your challenge
Level 3 · IndependentTake a process you were considering giving to an agent. Write two designs.
First, as a deterministic workflow: every step, every condition, and the point where a human is needed. Second, as an agent: goal, tool list, permissions, step cap, spend cap, and the approval gate for each irreversible action.
You have succeeded when you can state which is better for this specific process and give three concrete reasons — and when you have named at least one action your agent design must never take without a human.
Most people who do this exercise honestly discover the workflow wins.
What people usually get wrong
- Using an agent because agents are interesting. The interesting solution and the correct one are different questions.
- Granting broad permissions to save configuration time. Every extra tool is a wider blast radius on a bad day.
- No spend or step limit. A loop with no cap is a bill with no cap.
- Trusting the agent's report of its own success. Verify the world, not the summary.
- No log of what it actually did. When someone asks in March, an unauditable agent means nobody can answer.
- Letting it act on content it read from outside. Text that arrives from the internet or an inbox is data, never instruction.
- Skipping the boring baseline. Compare against the rules-based version. Often it is within a few percent of the agent, for a fraction of the cost.
How someone experienced does it
Experienced practitioners start with the agent to discover the process and then replace it with a workflow. The agent, watched closely on real cases, shows you which steps actually occur and where the exceptions are. Once you know, encode it. The agent was a research tool, not the product.
They also separate reading from acting, structurally. One component gathers and proposes; a different, dumber component executes only from a fixed set of approved actions. The intelligence is in the proposal, the safety is in the execution, and neither has to be compromised for the other.
And they measure the boring baseline before building anything. If a keyword rule gets 85% and the agent gets 91%, the honest question is whether the extra six points are worth the cost, the latency and the loss of explainability. Sometimes they clearly are. Usually nobody checks.
Why agents get more expensive as tasks get longer
Each step in an agent's loop typically includes the history so far, so the model can reason about what it has already tried. That means step 20 processes far more text than step 2.
Two consequences. Cost per step rises through a run, so a long run is disproportionately expensive rather than linearly so. And quality often falls as context grows — with a great deal of history, the important early instruction competes with everything since.
This is why long-running agents tend to drift from the original goal, and why breaking a large task into several short agent runs with clear handoffs usually outperforms one long autonomous run. It is also why "just let it run overnight" is rarely the good idea it sounds like.
Prove it
Take one process you believed needed an agent and design it as a workflow instead. Write down exactly where the workflow cannot cope.
Those points are the only places an agent is justified. Most people find there are one or two, in a process they were about to hand over entirely.
Keep learning this
Paste this into any AI assistant. It turns the assistant into a tutor that tests you instead of just answering you.
Act as an experienced practitioner who is good at teaching. I have just learned when to use an AI agent versus a deterministic workflow. Assume I am intelligent but relatively new to this — treat me as advanced level. Work through this in order, and wait for my reply at each step: 1. Ask me 5 questions that test whether I actually understood when to use an AI agent versus a deterministic workflow. Do not reveal the answers yet. 2. After I answer, tell me which parts I got right, which I got wrong, and which I only half-understand. Explain only what I misunderstood — do not re-teach what I already know. 3. Give me one practical challenge based on something I could genuinely encounter at work or in daily life. Do not solve it for me. 4. Evaluate my solution the way an experienced person would judge it, including what a professional would have done differently. 5. Tell me what to learn next, and why that comes next. 6. Give me trustworthy sources for deeper study — prefer official documentation, primary research or standards bodies over blogs and videos. Rules for you: no buzzwords. No motivational filler. Say "I'm not certain" when you are not certain, and tell me which parts of your answer I should verify myself. Clearly separate facts from your recommendations and your opinions.
Become independent at this
Use this when you want a path from where you are to actually good, with checkpoints you can test yourself against.
I want to become independently capable at deciding when AI agents are appropriate — not permanently dependent on AI, tutorials or step-by-step guides. Design a progression for me with five stages: Beginner, Guided practice, Independent practice, Real-world application, Professional level. For each stage tell me: - what I must know - what I must be able to do without help - the mistakes people make at this stage - one practical challenge - one real project that would prove I reached this stage - one way I can test myself honestly Then tell me the signals that I am ready to move to the next stage, and the signals that I have skipped ahead too early. Keep the theory to the minimum I actually need. Focus on ability I can transfer to situations you and I have not discussed.