Skip to main content

Agent Sprawl

Splitting work across multiple specialized agents — planner, executor, critic, validator — when a single agent with well-designed tools and prompts would do the job. Treating multi-agent as the default architecture rather than a last resort.

Cognition (makers of Devin) explicitly avoids multi-agent systems for this reason.

Why It Happens​

  • Multi-agent maps intuitively to human team structures
  • Framework demos make orchestration look easy
  • It feels more "sophisticated" and architecturally impressive
  • Genuine intuition that decomposition helps — just applied too eagerly
  • Research papers on multi-agent are exciting and plentiful

What Goes Wrong​

  • Telephone game — information degrades as context passes between agents, with each agent missing details
  • Conflicting decisions — parallel agents make inconsistent assumptions about architecture and style
  • Coordination overhead — orchestration, error handling, and retry logic across agents is complex
  • Higher cost — every agent hop is an additional LLM call
  • Harder to debug — failures span multiple agents with no single trace

What to Do Instead​

  • Start with a single agent — add complexity only when simpler solutions fail
  • Use tools, not agents — give one agent better tools rather than splitting into multiple agents
  • Read-only sub-agents only — if you must decompose, use sub-agents for information gathering only, not decision-making
  • Single decision-maker — keep one coherent entity making all consequential decisions

Multi-agent is warranted only when:

  • Domains are clearly distinct with separate tools and context
  • A single context window can't fit the required information
  • You need fundamentally separate conversation states

Signs You Have This​

  • You have agents whose only job is to pass data between other agents
  • A single well-prompted agent could do what your 5-agent system does
  • Agent coordination code is more complex than the actual task logic
  • You're debugging inter-agent communication failures more than actual task failures