LOOP guide · Episode 01
Hand it to the AI, or step in?
A one-page guide to deciding when a human stays in the loop. By Jose Nobile and Sebastián Figueroa.
In episode 1 we clashed over one idea. Sebastián said that “in the equation, I’m the dumbest part, by far,” and that there are more and more decisions he’d rather stay out of. Jose answered that no matter which frontier model he uses, he feels like he’s “talking to a junior.” Both are right, depending on the case. This guide turns that argument into a filter you can use today.
Step 1. Classify the task (30 seconds)
Answer yes or no:
- Can it be undone? If it goes wrong, can you fix it with a click, a rollback, or an apology email?
- Does it move money, customer data, or access? Payments, deleting files, credentials, sending data to third parties.
- Does it depend on context that only lives in your head? What you promised the client, who pays and who doesn’t, what’s going on in the business.
- Is there a verifiable “correct” answer? A test that passes, a number that reconciles, a format that’s met.
- What costs more: being wrong or being late?
Step 2. Put the task in one of three modes
| Mode | When | What you do |
|---|---|---|
| 🟢 Autopilot | Reversible and verifiable and no money or access involved | Let the AI decide. Spot-check the output. |
| 🟡 Copilot | Reversible, but needs your context or judgment | The AI proposes, you approve. Give it the context in writing. |
| 🔴 Human in charge | Irreversible, or touches money, data, or access, or a client and a promise are involved | You decide. The AI researches, drafts, and argues the other side. |
Short rule: if it’s irreversible, a human signs off. In the episode we said it about biology (“we’re playing with something that’s potentially irreversible”), but it applies just as well to a payment, a database, or an email to a client.
Step 3. Install brakes before you let go
In the episode we talked about agents that do far more than they were asked to, and bill for it. Before you let one run, check these brakes:
- Spend cap. A budget or usage limit in the tool, not in your trust.
- Quantity cap. “Max 50” in the configuration too, not just in the prompt.
- Logs. If something goes wrong, you need the evidence.
- Least privilege. Only the access the task needs. No “just give it access to everything, it’s easier.”
- Unplug button. Jose put it simply: “unplugging is still totally effective.” Know how to shut the agent down and revoke its keys.
- Isolation. One agent shouldn’t be able to talk to everything else.
Step 4. What AI doesn’t bring yet (and you do)
Jose listed the variables he decides with that AI can’t see:
- What did I commit to, and by when? The client wants it working on the date, not the perfect solution.
- Which client am I dealing with? They don’t all carry the same weight.
- How much budget (or how many tokens) do I have left for this and everything else?
- What’s the most cost-efficient thing I can ship now and improve later?
- What will I say if I miss the deadline? AI tends to overpromise (“it’s shameless”). You answer for what gets promised.
If your task depends on two or more of these, it belongs in 🟡 or 🔴.
Step 5. Move up a mode with evidence, not faith
A task can move from 🔴 to 🟡 and from 🟡 to 🟢, but only with a track record:
- Start in copilot and note every time you correct the AI.
- If several rounds in a row needed no important correction, try autopilot with spot-checks.
- If an expensive mistake shows up, drop back one mode. No drama.
Step 6. Write down your judgment
We closed the episode with Sebastián’s question: how do we give AI the context that lives in our heads today? Start small. For each 🟡 task, write 5 lines:
Task: What “done well” means: What it must NEVER do: Spend / quantity limit: Who it asks when in doubt:
Paste that at the top of your instructions. That’s your judgment, encoded.
Pocket rule
Reversible and verifiable → automate. Irreversible, involving money or a promise → a human in the loop.