Your goal
Build a bounded tool-using agent and explain its actual execution trace.
PROJECT LAB / MODULE 10
The setup planner: an agent you can inspect
Watch an agent choose a calculator, compare options and stop at its limit.
The situation
Two fictional setup options cost 3×18 + 2×12 and 4×20 + 1×15. The approved budget is 90. The agent may use a Calculator tool; it has no shopping or weather tool.
A first win
Calculate the first option as78 and the remaining budget as12.
Keep this artifact
An agent workflow, a tool-call trace and a three-case boundary report.
Explore the mechanism
This interactive model teaches the mechanism. It does not call a model, search your files or send messages.
Why this works
Tool choice is the distinguishing step
The model can choose a connected tool and use its result. A fixed chain has a predetermined model call instead. Merely writing the word agent in a prompt does not connect capabilities.
Tool output is evidence
The arithmetic result should come from the calculator call and be visible in the trace. A plausible narrated tool call is not proof that an execution happened.
A stopping rule makes failures inspectable
Set a maximum iteration count and an error path. If the task needs an unavailable tool, the agent should explain the missing capability rather than inventing a result.
Build it, step by step
Define tool authority
Write the task and whitelist: compare supplied costs, use Calculator, return a draft recommendation. It may not purchase, send or browse.
Check: The allowed actions and missing tools are explicit.
Need a hint?
All prices are fictional sample inputs.
Connect the real agent
Configure n8n AI Agent with a model that supports tool calling and attach the Calculator sub-node. Set Maximum Iterations to4 for this exercise.
Check: The node has at least one connected tool and a visible limit.
Need a hint?
A chat model working in a basic chain may not support tool calling.
Run both options
Supply the two cost formulas and budget90. Inspect the calculator input/result and the agent final answer.
Check: A is78, B is95; A fits and B exceeds budget by5.
Need a hint?
Verify the trace shows actual tool output, then check the arithmetic independently.
Inspect the trace
Record request, tool name, input, result, number of iterations and final answer. Mark which claims came from a tool and which from supplied data.
Check: A partner can replay why the recommendation was made.
Need a hint?
Do not treat hidden reasoning text as a reproducible tool log.
Test unsupported requests
Ask for current weather, an online purchase and a task that exceeds the configured limit.
Check: The agent reports a capability boundary or stops; no imaginary execution appears.
Need a hint?
Use your actual workflow error/stop route to demonstrate the limit.
Evaluate transfer
Change quantities and budget, rerun, and compare results against a manually computed answer.
Check: The agent adapts without a new system prompt.
Need a hint?
Save your own rubric for math correctness, tool use and unsupported actions.
Build with a clear contract
You may use only a connected Calculator tool. Compare optionA=3*18+2*12 and optionB=4*20+1*15 against a fictional budget90. Return each total, budget difference, tool evidence and a draft recommendation. Do not purchase, browse or send messages. If a tool is unavailable or the task requires another capability, say so. Stop within4 iterations. A human reviews the recommendation.
When it goes sideways
It answers math without a tool trace
The agent narrated a calculation rather than using the connected tool.
Try: Make exact arithmetic tool use part of the rubric and inspect execution data.
No tool call occurs
The model or configuration lacks tool-call support.
Try: Choose a documented supported model and verify the Calculator connection.
It loops or claims unavailable actions
Scope and iteration limits are missing.
Try: Set the limit, narrow authority and test an unavailable capability.
Review your evidence
Tick a criterion only after checking your own artifact. These are self-reported checks, not an automated certification.
Make it your own
Add an approved-note lookup
Add a read-only notes tool before the calculator. Keep its corpus small and require source IDs. Compare lookup failures with arithmetic failures.
- Only approved notes are reachable.
- Unknown facts remain unknown.
- Tool budget and error paths are demonstrated.
Check the mental model
What proves calculator use?
A weather question arrives with only Calculator connected.
Remember the distinction
Tool choice is the distinguishing step
The model can choose a connected tool and use its result. A fixed chain has a predetermined model call instead. Merely writing the word agent in a prompt does not connect capabilities.
Tool output is evidence
The arithmetic result should come from the calculator call and be visible in the trace. A plausible narrated tool call is not proof that an execution happened.
A stopping rule makes failures inspectable
Set a maximum iteration count and an error path. If the task needs an unavailable tool, the agent should explain the missing capability rather than inventing a result.
Go to the source
MODULE 10
n8n: AI Agent
Use the primary documentation to verify this part of your build.
Apply it in the project labMODULE 10
n8n: Supported tool-calling models
Use the primary documentation to verify this part of your build.
Apply it in the project labMODULE 10
n8n: Calculator tool
Use the primary documentation to verify this part of your build.
Apply it in the project lab