Last week you built the loop. This week you build the Act arrow. A tool is how a language model reaches something outside itself (a calculator, a database, a payment API), and it is simultaneously the moment your agent stops being a chat window and starts being able to do damage. Tools are capability plus risk surface, and this week is about designing that surface on purpose.
Week 4 called this step “Act” and left it as a black box. Here it is, opened:
Last week “Act” was a label on an arrow. You stepped through episodes where the agent “called Search” and a result appeared, and we deliberately did not ask how.
Five steps, four of which are ordinary software you write and test. Pages 01 and 02 take them one at a time.
Your Berlin supplier invoiced 4,817 units at €293 each, plus 18% VAT. What is the euro total? Before anyone tells you the answer, find out for yourself what a local model does with this question.
Everything here is copy-and-paste into one window. Start Ollama, then in a terminal run ollama run qwen2.5:0.5b once; after that you are just typing at a prompt. No files, no folders, no Python.
Produce: the filled table below, plus one sentence per round: what changed, and why you think it changed.
Paste the identical question five times and write down all five answers. Press ↑ to repeat the last one.
Nothing to interpret now: no wording, no percentage, just one multiplication. Does that fix it?
The model cannot run anything, so for this round you are the code that runs it. Send this, and it will reply with a request instead of an answer:
Write down the expression it asked for: is it the right formula? Then compute it yourself (phone calculator is fine; the correct value of 4817*293*1.18 is 1665429.58) and hand the result back:
Run the whole round three times. You have just done by hand exactly what the Python on page 03 does automatically.
| Run | Round 1 asked in words | Round 2 given the formula | Round 3 the expression it asked YOU to compute |
|---|---|---|---|
| 1 | |||
| 2 | |||
| 3 | |||
| 4 | |||
| 5 |
Round 1: asked in words. Five runs, five different answers, spanning five orders of magnitude, none correct (the answer is 1,665,429.58):
Run 4 is the interesting one: it wrote the right formula, 4817 * 293 * (1 + 0.18), and still produced a wrong number.
Round 2: handed the formula. Wrong every time, and now the wrong answers cluster near the right size, which is worse because they look believable: 999,999.8 · 1,286,483.4 · 1,240,908.4 · 115,085.4 · 1,194,414.8. So this was never a wording problem. Nothing in the model multiplies anything; it predicts the most plausible-looking continuation, and for long multiplication that is a number of roughly the right magnitude.
Round 3: you as the tool. In our runs the model produced a clean JSON request every time, and the expressions varied: (4817*293)+(4817*293*0.18) (correct) in one run, 4817*293 + 4817*0.18 (wrong) in others. After the result was handed back, it finished correctly every time.
So the two failures are different. Rounds 1–2 are a computation failure, and handing the arithmetic to something that computes removes it completely; in Round 3 that something was you. What remains is a planning failure: deciding which expression to compute. That one stays with the model, and it is what page 03 measures.
4817*293 + 18/100 in one run, 4817*293 + 18*293 in another. Arithmetic moved into the tool; planning stayed with the model, and that is what page 03 measures. Bigger models get this particular product right more often, which is also not the point: a token predictor is non-deterministic and unauditable: you cannot promise a controller that the number will be right next month, and you cannot show the work. A tool call gives you exactness, repeatability, and a trace: three things a finance team will ask for.Press Play and watch who does what. The model only writes text shaped like a call. The harness (your Python loop) validates it and runs the function. The human gate sits in front of the one action that cannot be undone. Switch scenarios to see an invalid call repaired, a gate decision, and a small model that cannot recover.
Try this: step with the → key. At the gate choose Reject, then replay and choose Approve. Which single line in a trace log would let an auditor tell the two runs apart a month later?
A model without tools is a very expensive autocomplete. Every tool you add is a verb the agent acquires: read, compute, file, send, pay. The business value of an agent is almost entirely the value of its verbs.
Every verb is also a way to be wrong at scale. A read tool leaks; a write tool destroys; a spend tool costs money. The schema you write is where “what is this agent permitted to do?” gets answered: in code, not in a policy document.
Because every tool call is a structured record (name, arguments, result, timestamp), a tool-using agent is far easier to audit than one that “just knows.” When compliance asks what happened, you replay the calls.
Background, not a prerequisite. You can do pages 01–03 without it. It answers the question one step behind this week: why give a model tools at all?
A model's weights were frozen on a date. It cannot know your Q3 numbers, today's outage, or a policy you published last Tuesday. Closed by: a retrieval or search tool.
Your CRM, your ticket queue, your inventory table were never in anyone's training data, and should not have been. Closed by: a read tool against your own systems.
Predicting the next token is not arithmetic, and it is certainly not running a SQL query or a pricing model. Closed by: a calculator, a code runner, a query tool.
Text cannot file a ticket, send an email, or issue a refund. Closed by: write tools, and this is exactly where the risk arrives.