Calling the tool isn't the finish line. You have to read the result that comes back, understand it, and only then decide the next step.
Let's think about this for a second
Most agent mistakes come from guessing the answer instead of actually reading the tool result. It'll claim something is "in there" when it isn't. So set a rule: every fact in the answer has to be traceable back to the latest tool result. If it isn't, the answer should say "not found in the notes." Quoting a short snippet back lets a human verify it.
Let's connect this to everyday life
Say read_note("photosynthesis.md") returns "Plants use sunlight to make food." A good answer quotes this sentence and explains it. A bad answer goes off and invents chlorophyll's chemical formula. In the lesson exercise, you'll be given a tool result and practice telling a correct answer apart from a wrong one. While you're coding, print the tool result to the log so you can actually see it.
Let's try it out together
Tool result:
photosynthesis.md → အပင်များသည် နေရောင်ကိုသုံး၍ အစာလုပ်သည်
Good answer:
မှတ်စုအရ အပင်က နေရောင်သုံးပြီး အစာလုပ်ပါတယ်။
Bad answer:
Chlorophyll က CO2 နဲ့ H2O ကို C6H12O6 ပြောင်းပါတယ်။ (ဖိုင်ထဲမရှိ)You'll be able to tell an answer that's grounded in the tool result apart from one that's guessed.5-minute try
Write a short tool result. Then write out both a correct agent answer and a guessed one, and explain the difference.
A quick word of caution
Sounding confident isn't the same as being correct. An agent can state something firmly and still not match the tool result.