Week 6 Reflection: Custom Agents¶
Purpose¶
This week was about moving from "the agent" to designed agents.
A custom agent is not just a different prompt. It is a role with a model, mode, description, and permission boundary.
Use this reflection to check whether you can design agents intentionally.
Short Answers¶
Write 3–5 sentences for each question.
1. Primary vs. Subagent¶
Explain the difference between a primary agent and a subagent in your own words.
Then classify the two agents you created:
| Agent | Primary or Subagent? | Why? |
|---|---|---|
| Agent 1 | ||
| Agent 2 |
2. Agent Anatomy¶
Pick one of your agents and fill this out:
| Field | Your Agent's Value | Why It Matters |
|---|---|---|
| Mode | ||
| Model | ||
| Prompt | ||
| Description | ||
| Permissions |
If any field feels vague, revise that part of the agent before moving on.
3. Permissions¶
What is one tool or action your agent is not allowed to use?
Why is that restriction useful?
Be concrete. For example:
The docs writer cannot run bash because documentation work does not require command execution, and shell access could let it run risky or irrelevant commands.
4. Boundary Test¶
Describe one forbidden-action test you ran in Lab 6.2.
What did you ask the agent to do?
What did you expect?
What actually happened?
Did the result prove a prompt boundary, a permission boundary, or both?
5. Agent vs. Command¶
Think back to Week 4.
Name one task that should stay a command instead of becoming a custom agent.
Name one task that deserves a custom agent.
Explain the difference.
Use this rule:
- Command: reusable prompt, same role, same permissions
- Agent: different role, model, or permissions
Design Review¶
Review your two agents as if you were approving them for a team.
| Question | Agent 1 | Agent 2 |
|---|---|---|
| Is the role clear? | Yes/No | Yes/No |
| Is the mode correct? | Yes/No | Yes/No |
| Are permissions minimal? | Yes/No | Yes/No |
| Is the description specific? | Yes/No | Yes/No |
| Would you let a teammate use it? | Yes/No | Yes/No |
For every "No," write one concrete change.
Exit Ticket¶
Before moving to Week 7, you should be able to say yes to all six:
- [ ] I can create an agent with
opencode agent create. - [ ] I can explain primary vs. subagent.
- [ ] I can identify an agent's mode, model, prompt, description, and permissions.
- [ ] I can choose permissions based on the agent's job.
- [ ] I can test whether a forbidden action is actually blocked.
- [ ] I can decide when to create an agent instead of a command.
Looking Ahead¶
Next week, you will connect agents into workflows.
The pattern will be:
Primary agent decides what needs to happen.
Subagent does focused work in its own context.
Primary agent reads the report and chooses the next step.
That only works if this week's agents have clear roles and safe boundaries. If your agents are vague now, multi-agent workflows will become noisy later.