Module 1 · Foundations of Agentic AI · scripted

Tool Parameters and Descriptions

20 minoutcomes: tool-design

Tools have controls

Declaring which tools exist bounds what actions the agent can take. But a real tool also constrains how it may be used. A microwave doesn't accept "warm it gently" — it accepts a power level and a time. An oven wants a temperature, a rack position, and which heating elements are on: top, bottom, or both.

These are the tool's parameters: the inputs a tool requires before it can act at all.

Two tools, spelled out

MICROWAVE
power: level 1–5
time:seconds
OVEN
temperature: °C, 50–250
rack:1 (bottom) – 4 (top)
elements:top | bottom | both

Say exactly what the tool accepts

Here is where precision starts to matter enormously. How a parameter must be expressed depends on the actual tool — and the LLM can only know what the conversation tells it:

The same microwave, four honest descriptions

A. power:integer 1–5 → "power 3" ✓ "power 60%" ✗
B. power:percent in steps of 10 → "power 60%" ✓ "power 3" ✗
C. power:any decimal 0–5 → "power 1.6732" ✓
D. power:integer 1–3 → "power 3" ✓ "power 4" ✗
Same appliance, different contract
  • Why is "power 60%" a perfectly good answer for microwave B and a meaningless one for microwave A?
  • What happens to the conversation if the LLM starts issuing "power 60%" commands at microwave A?

Why be this careful? Because of everything we know about trajectories. At the start of the conversation, we want to provide enough information that the LLM heads off on a trajectory that leads efficiently to a correct solution. If it begins specifying power levels in a format the tool doesn't accept, those invalid actions are now in the conversation — conditioning what comes next, costing correction turns, and bending the trajectory. (There are ways to catch and repair invalid tool use, which come later — but repair is always the expensive path. Precision up front is the cheap one.)

Example 1 · The contract at work
user

Reheat my leftover fiskesuppe. Tools: MICROWAVE (power: integer 1–5; time: seconds). It's a creamy soup, so be gentle — high power splits the cream.

assistant

Gentle it is: microwave at power 2 for 90 seconds, then stir and tell me how it feels — we'd rather sneak up on hot than split the cream.

user

Stirred. Warm at the edges, cold in the middle.

assistant

Classic microwave cold spot. Power 2 for another 60 seconds, stir again. If it's still not hot after this round, we'll go to power 3 in 30-second bursts.

NoteEvery command is legal — integer power, seconds — because the contract was in the conversation before the first action. Nothing needs repairing.

Tools the LLM has never heard of

There is a second thing every tool may need: a description. Not all tools are well known to the LLM. It has never seen your lab's equipment, your company's internal software — or the gadget you bought from a TV offer last month, which did not exist when the model was trained:

Example 2 · The gadget that post-dates the model
user

New tool in my kitchen. GRILLVENN G-1: a countertop gadget that clamps around a fish fillet and cooks it from both sides at once between two ribbed ceramic plates. Parameters — plates_temp: LOW | MED | HIGH; clamp_time: whole minutes, 1–10. It beeps when done and must rest closed for 1 minute after beeping before opening. Now: cook my salmon fillet with it.

assistant

Understood — the GRILLVENN does the whole job, both sides at once, so there's no flipping step. For a standard fillet: plates_temp MED, clamp_time 4 minutes. When it beeps, don't open it — let it rest closed for its 1 minute so the juices settle. Then open and check the center; if it needs more, we'll clamp again at LOW for 1 minute rather than re-firing at MED.

NoteThe model has never seen a GRILLVENN — it couldn't have; we invented it after training. Everything it "knows" comes from the description in the conversation: no flipping (both sides at once), legal temperatures, whole-minute times, the rest-after-beep rule. And its plan honors every one of those constraints.
Where does tool knowledge live?
  • List three things the agent "knew" about the GRILLVENN. Point to the words in the conversation each one came from.
  • What would the agent have done with "GRILLVENN G-1" and no description?
  • Which of your own domain's tools would an LLM genuinely know — and which only feel well-known to you?

So a tool, fully stated, has three parts: a name, a description of what it does and how it behaves, and parameters with exactly what values they accept. Get those three right in the conversation, and the agent's actions arrive legal, precise, and executable — which is exactly what we'll need when the participant performing the actions stops being a patient human and becomes a computer.

Your turn

Take one real tool from your research — an instrument, a dataset query, a piece of software. Write its name, description, and parameters precisely enough that an LLM could operate it without a single invalid action. Then test it: give the spec and a goal to an LLM, and audit every action it proposes against your spec.