Exercise 4: Subagents¶
Duration: 20 min | Module: 4 — Multi-Agent & Advanced
Objective¶
Spawn parallel subagents on your codebase, practice the adversarial reviewer pattern, and observe isolated execution.
Part 1: Parallel Audit (10 min)¶
Launch agy interactively:
Dispatch a parallel audit team:
> Spawn two subagents in parallel using branch workspace mode:
> 1. A security auditor — scan for: hardcoded credentials, injection vulnerabilities, exposed sensitive data, and insecure dependencies
> 2. A test coverage auditor — identify: untested public functions, missing edge cases, and integration test gaps
>
> Report back when both complete with a combined findings summary.
While they run, ask:
When they finish:
Part 2: Adversarial Reviewer (7 min)¶
Pick a recent PR, branch, or any set of changes:
Back in agy:
> I have changes on the current branch. Spawn an adversarial reviewer subagent.
> Its only job: find reasons why these changes should NOT be merged.
> It should challenge assumptions, look for edge cases, and be skeptical of everything.
> Be harsh — this is an adversarial review, not a supportive one.
Read the adversarial findings. The goal is to identify what a thorough code review would catch.
Part 3: Resume a Subagent's Work (3 min)¶
> One of the subagent findings mentioned [specific issue]. Let's fix it. Create a subagent in inherit mode to implement the fix.
Note the difference from branch mode: inherit means the subagent works in the same directory as your main session — appropriate for targeted, non-conflicting fixes.
Completion Criteria¶
- [ ] Spawned at least 2 parallel subagents successfully
- [ ] Both subagents ran and returned findings
- [ ] Adversarial reviewer returned critical findings
- [ ] Used at least two different workspace modes (branch vs inherit)