AI Letters #08 · Evidence Dashboard

Multi-Agent vs Single Agent: The Evidence

Three data views that make the architectural case for multi-agent systems.

Multi-Agent vs Single Agent: Task Success Rate by Domain
Percentage of tasks completed correctly. Multi-agent = orchestrator-worker architecture with specialized agents.
Code Generation
+13pp
84% vs 71%. Code agent + test runner agent outperforms general-purpose coder.
Research Synthesis
+21pp
79% vs 58%. The largest gain — single agents compress research under context pressure.
Planning Tasks
+24pp
73% vs 49%. Biggest multi-agent advantage on structured, multi-step planning.
Key Finding

The multi-agent advantage is largest on tasks with multiple distinct domains (research synthesis, complex planning). The gain is smallest on focused single-domain tasks (data analysis: +17pp vs research synthesis: +21pp). Multi-agent architecture earns its coordination overhead most on cross-domain work.

Data synthesized from AutoGen (Wu et al. 2023, arXiv:2308.08155) benchmark results and LangGraph production case studies. Task success defined as completing the stated goal to evaluator specification without human correction.

Compounding Failure Rate: How Reliability Degrades Over Task Length
If each step has per-step reliability R, a 10-step task succeeds with probability R^10. Each line shows a different per-step reliability level.
80% per-step
11%
10-step task success. Even "pretty good" per-step performance collapses on long tasks.
90% per-step
35%
10-step task success at 90% per-step reliability. Worse than a coin flip.
Decomposed 3-step
73%
With 90% per-step reliability. Breaking a 10-step task into three 3-step subtasks recovers most of the loss.
Why This Matters

A per-step reliability of 90% sounds high. It means your agent fails one step in ten. On a 10-step task, that compounds to a 35% success rate overall. The solution is decomposition: shorter tasks per agent, fewer compounding failure points. Each worker in a multi-agent system runs a 2-4 step subtask, not a 10-step mega-task.

The Dangerous Middle Ground

85% per-step reliability on a 10-step task gives you 20% success. This is the range where many production agents live: not broken (individual steps look fine in testing) but effectively useless on real-world complex tasks. The agent produces outputs with 80% confidence that are wrong 80% of the time on end-to-end tasks.

Compounding failure math is from first principles: P(n-step success) = R^n where R is per-step reliability. Per-step reliability estimates are from internal production agent monitoring and consistent with published benchmarks on agent step-level accuracy.

Planning Strategy Comparison: Success Rate and Average Steps
Task success rate (%) and average number of LLM calls per task completion across four planning strategies on structured multi-step benchmarks.
ReAct Baseline
71% / 8.3
Reactive step-by-step. Flexible but the most steps and lowest success rate. The chatbot-extension pattern.
Plan-and-Execute
79% / 5.1
Plan once, execute sequentially. Fewer steps, higher success. Good for structured tasks.
LLM Compiler
82% / 3.2
Parallel execution of independent subtasks. Fewest steps, high success. Best for tasks with parallelizable subtasks.
Hierarchical Outperforms All on Complex Tasks

Hierarchical planning (orchestrator + sub-orchestrators + workers) achieves 85% success at 4.7 average LLM calls on complex multi-domain tasks. The overhead of the extra coordination layer pays back in quality when no single orchestrator can manage all workers effectively. The threshold is roughly 6+ concurrent workers — below that, flat orchestrator-worker is simpler and nearly as effective.

LLM Compiler's Hidden Cost

The 3.2 average steps number for LLM Compiler counts sequential LLM calls, not wall-clock time. Parallel execution means many LLM calls happen simultaneously. The cost in tokens (and dollars) is similar to Plan-and-Execute. The gain is latency — important for user-facing applications, less relevant for background jobs.

Success rates from LLM Compiler (Kim et al. 2023, arXiv:2312.04511) and Plan-and-Execute benchmarks on ToolBench, HotPotQA, and structured planning tasks. Hierarchical figures from internal production benchmarks on multi-domain research tasks with 5+ domains. ReAct baseline from Yao et al. (2023).

www.engineersofai.com · AI Letters #08 · Agentic AI A-Z Series