Claude Certified Architect – Foundations Practice Question 40: Long Horizon Progress Notes

Claude Certified Architect, Foundations · Test 4 · question 10 of 10

Context Management & Reliabilityhard

A team runs an agent that builds a large application over many hours, across many separate context windows. Each new session starts with no memory of the last, and the agent tends to redo finished work, declare the project done early, or leave the code half-broken. Which harness design best addresses this?

Select one answer
Show answer and explanation

Answer:

  • A. Have each session write its progress to files in the workspace (a feature checklist with pass/fail status, a progress log and git commits), and have the next session read them first and work on one item at a time.

Anthropic's guidance for long-running agents is to bridge context windows with external artifacts: a structured feature list, a progress file and version control, so each new session can quickly see what is done, what is next and what state the code is in. Working one feature at a time and leaving a clean state reduces half-finished work. A memory tool serves the same purpose of persisting notes outside the context window.

B fails because even the largest window eventually fills and costs grow. C bloats the context with irrelevant history and eventually exceeds the limit. D cannot work: instructions do not create persistence across sessions.

Reference: Effective harnesses for long-running agents

Practice Test 4 ›

Study guides for Claude Certified Architect, Foundations