🤖 What Is an Agent
A chatbot answers a message and stops. An agent does more: it reasons about a goal, uses tools (search, code, APIs, your calendar) and completes the task in several steps — deciding the next step based on the previous result. That’s the difference between someone who answers a question and someone who solves the problem.
The agent loop: goal → plan → tool → observe → repeat until complete — illustrative diagram.
🥤 Vending Machine × Slot Machine
The best analogy: a workflow is a vending machine — press B4, get the same can, every time. A agent is a slot machines — same input, different result each time. Both are useful; just don’t confuse which one you’re using. For most repeatable work, you want the predictability of the vending machine.
✓ Vending machine (workflow)
- ✓Deterministic: same input → same output.
- ✓Cheap, fast, and easy to audit.
- ✓Ideal for repeatable tasks: "every day at 9 a.m., pull and post".
- ✓When it breaks, you know exactly where.
✗ Slot machine (agent)
- ✗Nondeterministic: results vary each time.
- ✗More expensive and slower per run.
- ✗Difficult to predict and audit step by step.
- ✗Only worthwhile when there’s genuine ambiguity.
🔧 When You Need AI × When You Don’t
The right question isn’t “how do I put AI here?” but “is this task deterministic or ambiguous?”. If every step is fixed and predictable, a simple workflow—cheap and robust—will do the job. If the task requires judgment and reasoning on top of inputs that change, that’s when the agent earns its place.
🧪 The one-question test
"Could I write the exact rule in a flowchart?" If yes, it’s a workflow: "new lead in the CRM → send a welcome email". If no, because it depends on interpreting free-form text, deciding case by case, or researching before responding, so it's a candidate for an agent. Always start with the cheapest option that works.
🧭 Intent Is What Matters
Agent or workflow, what sets your result apart isn’t the technology—it’s the intent architecture. You stop sending messages and start designing the system: what context goes in, what rules apply, what goal to pursue, and how the result is validated before moving forward. The phrase is disposable; the intention you encode is the asset.
🧭 The four pillars of intent
- •Context: what the system knows about you, the client, and the world.
- •Rules: what it can do, what it can’t do, and when to stop.
- •Objectives: the definition of "done" — the metric that counts.
- •Validation: how you check it before it reaches the client.
⚠️ Risk increases when you’re out of the loop
While you review every output, an agent’s mistake lands on your desk. The moment it runs without you in the loop — on its own, in the background — the same mistake goes straight to the customer. Removing the human from the loop multiplies leverage, but also multiplies risk. The more autonomous it is, the more the system needs to be foolproof.
⚠️ Alert: out of the loop, the error goes public
Before letting an agent run unsupervised, cover the security basics:
- ▸Hard limits: what it must never touch (delete data, send money, talk to a client without review).
- ▸Automatic validation: checks that block out-of-pattern output before publishing.
- ▸Stopping point (human-in-the-loop): high-risk actions wait for your OK.
- ▸Stress test: run it with intentionally bad inputs and see whether it fails safely.
🏅 The elite move: “AI isn’t needed here”
Beginners cover all of AI to seem advanced. The expert looks at the problem and says: "this is a formula, a filter, three lines of code—it doesn’t need AI". Solve more cheap, more fast and more safe is far more impressive than putting an agent in every little corner.
💡 Practical tip
Before designing your dream agent, ask: "what’s the simplest solution that delivers the same result?" Often it’s a spreadsheet, a no-code automation, or a fixed rule. Save the agent for what truly requires reasoning—and build a reputation as someone who solves problems, not complicates them.
Self-recovery (optional): when should you NOT use an AI agent?
✅ Module Summary
Next Module:
3.4 — Build Your Jarvis: always-on partners