Multi-agent made it worse
“A team split a single agent into a supervisor with four workers and quality dropped. What are the likely causes and what would you do?”
What this tests
- Diagnosing information loss and coordination failures from traces
- Understanding the supervisor as a bottleneck and a lossy summariser
- Willingness to revert or simplify based on evidence
- Knowledge of agent-to-agent communication design
Answers by level
Read the beginner answer first and notice what is missing.
The usual causes: the supervisor loses information when summarising the task for workers and their results for itself; workers lack context the single agent had (the user's original phrasing, earlier tool results); hand-offs use free-form prose so details drift; the supervisor makes extra round trips and hits iteration or budget caps; and errors compound because a worker's mistake is taken as truth by the supervisor. See Supervisor Pattern and Agent-to-Agent Communication.
I would pull traces for the regressions and compare the single-agent run with the multi-agent run on the same input: where did the needed fact disappear? Then fix the boundary: pass structured messages including the original request, make workers return typed outputs with confidence and citations, cap supervisor rounds, and give the supervisor the raw worker outputs rather than its own summaries when they fit the budget.
If after that the multi-agent version is not measurably better on the golden set, revert. The single agent was a working baseline; the split has to earn its place.
Green flags · Red flags
- Names information loss at hand-offs as the primary suspect
- Compares traces of both versions on the same inputs
- Proposes typed messages that include the original request
- Considers converting to a pipeline or merging workers
- Willing to revert on evidence
- Adds per-hop evaluation
- Tunes prompts blindly
- Adds more agents to fix a multi-agent problem
- No trace comparison or measurement
- Treats reverting as unacceptable