← All courses133 · Study guide

Agent harnesses for long-running tasks

Agent designApp and web development

What you will be able to do

  • Track task state, retries, handoffs and evidence of completion. Resume safely after interruption without repeating completed side effects.

Core concept

  • A model may request an action; an executor validates and performs it. Observe tool results before choosing another action. Success, a tool error and reaching an action limit are different states.

Practice 1 · Try a small task

  • Sketch a two-step agent for a synthetic price quote: look up a unit price of 5,000 KRW, then calculate two units. Define each input, output and a two-action limit. Do not connect a real order or payment system.

Practice 2 · Failure and verification

  • Make the lookup return NOT_FOUND, then try an action limit of one. Write the expected stop state in each case. Add a human approval boundary before any real sending, purchasing or deletion.

Review your work

  • Check the task contract, minimum permissions, observed tool results and final evidence. An agent saying it finished is not the same as a verified result.

Use synthetic or anonymized material for practice. A person must review before real sending, payment or deletion.

These guide exercises do not award completion, XP or badges. Complete the existing verified labs and quiz in the full lesson to earn them.

The full lesson requires sign-in and course access.

Full lesson in Korean →
133 Agent harnesses for long-running tasks · Study guide · AIpiens Academy