Skip to content

Week 7 Reflection: Boundaries Make Agents Useful

Purpose

This reflection helps you decide when multi-agent workflows are worth using and how to keep them safe.

Answer in plain language. Short, specific answers are better than long general ones.

Part 1: Your Mental Model

  1. In your own words, what is the job of a primary orchestrator agent?
  2. In your own words, what is the job of a subagent?
  3. What makes a subagent different from simply continuing the same conversation with the primary agent?
  4. Why can context isolation be useful?
  5. Why can context isolation also be inconvenient?

Part 2: Hand-Off Quality

Review one prompt you wrote this week to delegate work to another agent.

Paste the prompt, then answer:

  1. What role did you assign?
  2. What scope did you define?
  3. What did you explicitly tell the agent not to do?
  4. What output shape did you request?
  5. If you ran the same workflow again, what would you tighten?

Part 3: Failure Modes

Choose one failure mode from the lesson:

  • Context bleed
  • Infinite loop
  • Conflicting edits
  • Noisy hand-off

Answer:

  1. What does this failure look like in practice?
  2. What is one early warning sign?
  3. What is one prompt or permission change that would reduce the risk?

Part 4: Judgment

For each task, decide whether you would use one agent or multiple agents. Explain why.

  1. Fix a typo in one Markdown file.
  2. Review a pull request for security and style issues.
  3. Rename a function across a small project.
  4. Investigate a flaky test in a large codebase.
  5. Write release notes from a clean commit history.

Part 5: Your Rule Of Thumb

Complete these sentences:

I should use one agent when...

I should use multiple agents when...

Before I let more than one agent edit files, I should...

The most dangerous multi-agent mistake for me is probably...

Submission Checklist

Before you submit, check that your reflection includes:

  • One clear definition of primary orchestrator
  • One clear definition of subagent
  • One reviewed hand-off prompt
  • One failure mode explained with a prevention tactic
  • A personal rule of thumb for future work