Review: Designing Real Agents
Before the capstone.
Before you build Juno, let's check the design ideas from this last stretch. These are the calls that you'll actually have to make at work. Pause, and commit to an answer each time.

- Check your design instincts
- Calls you'll make at work
- Commit each time
Recap: design
Show the plan, ask clearly, and make actions reversible. Voice has a budget of about one second. Background agents prepare drafts. Safety comes from permissions, rather than instructions. Keep long plans on disk. Compact before you hit the limit. And the user's identity never passes through the model.
Question 1
Question 1. Which approval is best? A, Approve action?. B, Refund $640 to Priya for damage, reversible for 24 hours. C, Yes or no. D, Continue?. Pause and choose. The answer is B. People can make a real decision.
Question 2
Question 2. What keeps a voice agent feeling fast? A, Longer replies. B, Streaming, and speaking immediately. C, Bigger models. D, Slower speech. Pause and choose. The answer is B. Fill the silence.
Question 3
Question 3. What should a new background agent do? A, Send emails directly. B, Prepare drafts for review. C, Delete old data. D, Nothing. Pause and choose. The answer is B. Earn trust first.
Question 4
Question 4. What's the real wall for a SQL agent? A, A polite prompt. B, A read-only database user. C, A long description. D, A strict tone. Pause and choose. The answer is B. Make damage impossible.
Question 5
Question 5. How does an agent keep working across sessions? A, A longer context. B, Plans and notes in files. C, Higher temperature. D, More tools. Pause and choose. The answer is B. State on disk beats memory.
Question 6
Question 6. Who should supply the user ID to a tool? A, The model. B, The session, outside the model. C, The customer. D, Nobody. Pause and choose. The answer is B. Otherwise an attacker just asks for it.
Summary
Next, the lessons of real failures.
Key takeaways- Readable approvals
- Permissions over prompts
- Identity outside the model