Question 18
Domain 2A team spawns subagents for 'cleanliness' on a sequential task: extract → transform → load. Each step depends on the previous. What's the structural critique?
Correct answer: C
Explanation
Sequential work has no parallelism, so spawning separate subagents adds overhead without benefit. Because each step depends on the previous, there are no isolation triggers to justify splitting the task, and “each subagent is a full model invocation chain” means extra token cost. A single inline loop fits the dependency structure better.
Why each option is right or wrong
A. Each subagent should pass state via shared memory.
Shared memory helps coordination, but it does not justify splitting a purely sequential pipeline.
B. The subagents should run in parallel anyway and reconcile at the end.
Dependent ETL stages cannot meaningfully run in parallel before prior outputs exist.
C. Sequential work has neither parallelism nor isolation triggers — over-spawning burns tokens (each subagent is a full model invocation chain) without structural benefit. Use a single inline loop.
The dependency chain here is strictly linear: extract must finish before transform can begin, and transform must finish before load can run, so there is no concurrency to exploit. In the absence of parallelizable branches or an isolation boundary, spawning separate subagents adds only orchestration overhead; each subagent is a full model invocation chain, so the design burns extra tokens without improving correctness or throughput. A single inline loop matches the sequential control flow and avoids unnecessary agent fan-out.
D. The middle subagent (transform) should be replaced with a tool.
A tool is for deterministic capability, not a structural fix for unnecessary agent decomposition.