An LLM can only talk. How does it learn to do things?
On its own, a language model just predicts text, it can't check the weather, run a calculation, or search a database. An agent is what you get when you wrap that model in a loop: it thinks up a plan, acts by calling a tool, observes the real result, and repeats until the task is done. Below is a real agent loop you can single-step. Turn its calculator off and watch it try to do math in its head. Then, in Part 2, run 50 of them at once and see what breaks.
▶ Part 1 · One agent: pick tools, set a step limit, then Run
📋 The task (fixed)What's an 18% tip on an $86.40 bill, and is the total over $100?
Tools the agent can call:
2 · tightgenerous · 8
step 0 / 6tools used 0status idle
The loop hasn't started. Press Run to auto-step, or Step to advance one stage at a time.
▶ Part 2 · Now run many agents at once
Agents are cheap to copy, so teams run many in parallel. Every result an agent hands back still needs a human to read it, understand it and approve it. People have a fixed number of hours in a day. What happens when the agents produce more than people can check?
🔮 Commit a guess first
You go from 1 agent to 20. Each result needs 15 minutes of human review, and the team has 4 review hours a day. After two weeks, the review queue holds…
Pick one. Then drag Parallel agents to 20 and check.
150
120
1 min60 min
1 h8 h
0%95%
Review load38%of human review time
Queue grows by0results per day
Queue after 2 weeks0results waiting
Comprehension debt0 hof review nobody has done
🧭 Where this fits
The agent loop, model decides, tool runs, result feeds back, is the skeleton of every AI agent shipping today: coding assistants, browser agents, customer-support bots. The two knobs you just played with (which tools, how many steps) are exactly what production teams tune, and the step cap is the difference between an agent and a runaway bill. Part 2 is the other half of running agents for real: human attention is the scarce resource. Automated checks (tests, type checks, evals) are the leash that lets you run more agents without drowning in review.
The code that's actually running
🔮 Predict, then test
Turn the 🔧 Calculator OFF and Run, does the agent still get the exact total, or does it confidently guess? Why does giving a model tools beat asking it to "just think harder"? Now set max steps to 2 on this task (it needs three actions to finish) and Run, what stops it, and why is that stop condition the difference between an agent and an infinite loop?
The takeaway
Your new mental model of an agent
An agent = LLM + tools + a loop: think → act → observe → repeat. The model supplies the reasoning; the loop supplies the doing.
Tools ground the model in reality, a calculator, a search index, a code runner return facts, so the agent stops guessing at math and quotes the real result.
The loop lets it decompose a task and react: each observation becomes context for the next thought, so a multi-step job gets solved one grounded step at a time.
Without a stop condition, a "done" signal or a max-steps guard, an agent can loop forever or burn your budget. The limit isn't a nicety; it's a safety rail.
Agents scale, human attention doesn't. When results arrive faster than people can review them, the queue grows every single day. It never catches up by itself.
The unread pile is comprehension debt: work that exists but that no one on the team understands yet. Merging it unread doesn't pay the debt, it hides it.
Verification is the leash. Tests and automated checks that fully verify a share of results let you run more agents before people become the bottleneck.