The Load Balancer Trap in Multi-Agent Systems
TRIGGER
When orchestrating multiple subagents, vague task descriptions like 'research the semiconductor shortage' caused agents to duplicate work, leave gaps, or misinterpret tasks—one subagent explored 2021 automotive chip crisis while two others duplicated work on 2025 supply chains.
APPROACH
Anthropic changed their lead agent prompts to require detailed task descriptions for each subagent. Each delegation now includes: (1) specific objective, (2) output format, (3) guidance on tools and sources to use, and (4) clear task boundaries distinguishing this subagent's scope from others. They also embedded scaling rules: simple fact-finding gets 1 agent with 3-10 tool calls, direct comparisons get 2-4 subagents with 10-15 calls each, complex research gets 10+ subagents with clearly divided responsibilities. Input: user query complexity assessment. Output: structured subagent task assignments with explicit boundaries.
PATTERN
“Two agents duplicating research on 2025 supply chains while a third explores 2021—the "load balancer trap" where vague delegation creates overlap and gaps. The fix isn't more agents; it's explicit scope boundaries. Each subagent needs: specific objective, output format, tools to use, AND what it explicitly doesn't cover.”
✓ WORKS WHEN
- Query can be decomposed into 2+ non-overlapping subtasks
- Subtasks have natural boundaries (different time periods, different entities, different source types)
- Lead agent has enough context to understand task structure before delegation
- Output format can be specified clearly enough that results are mergeable
- Subagents don't need to coordinate with each other during execution
✗ FAILS WHEN
- Task boundaries are unclear or heavily overlapping in reality
- Subtasks have dependencies that require inter-agent communication
- Query is too novel for orchestrator to understand decomposition structure
- Subagent results need real-time synthesis rather than post-hoc merging
- Domain requires shared context that can't be replicated across subagents