Does coordination improve correctness?
Compare single-agent and multi-agent workflows on identical tasks, with independent verification and explicit failure categories.
Experimental conditions ↗ INDEPENDENT RESEARCH INITIATIVE / 2026
NOAERTH SCHOLAR RESEARCH NOTE 001
When should AI agents work together?
Multi-agent systems promise greater capability. They also introduce coordination costs, hidden failures and difficult trade-offs. We are designing controlled experiments to find where collaboration actually helps—and where simpler systems are better.
Current status: protocol design. No original empirical results or peer-reviewed publications are claimed.
FIGURE 01CONCEPTUAL SYSTEM
POLICY
+ OVERSIGHTAGENT AAGENT BAGENT CEVALUATOR
+ OVERSIGHTAGENT AAGENT BAGENT CEVALUATOR
Illustrative topology · Not measured output01 — 04
RESEARCH DOMAINAutonomous AI systems
FOCUSCorrectness / cost / oversight
EVIDENCEProtocol under development
INDEPENDENCENo university affiliation
01 / THE RESEARCH
A narrower question.
A more useful answer.
Claims about agents are easy to make. Reliable comparisons demand shared tasks, transparent evaluation and an account of the resources spent getting to a correct answer.
What does a correct result actually cost?
Measure full-system expenditure, retries, latency and cost per verified completion—not just headline model prices.
Evaluation metrics ↗When should humans intervene?
Test whether approval gates reduce material errors and whether intervention costs outweigh the benefit.
Risks and limitations ↗02 / METHOD
Not another
performance claim.
We will publish results only after trials can be traced to task specifications, model versions, execution records and independent scoring.
01Fix the questionPredefine primary outcomes, task strata and evaluation rules.
02Compare credible baselinesSame tasks, compute budget and environment across workflows.
03Expose negative evidencePublish coordination overhead, failures and non-improvements.
04Make repetition possibleVersion task sets, model choices, prompts and result artifacts.
PROTOCOL STATUS● IN DESIGN
hypothesisheterogeneous vs. single-agent
controlssame tasks, scoring, budget
metricsverified success / cost / risk
resultsNOT YET AVAILABLE
confidenceNOT ESTIMATED
Visualizes planned fields only. Not a live experiment feed.
03 / SCHOLARSHIP
Existing literature
shapes the study.
This project starts from prior work, including multi-agent orchestration, model cascades and agent benchmarks. The goal is to determine what still requires rigorous testing.
01 · SYSTEMS
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
Wu et al. · 2023 / COLM 2024 · Multi-agent composition and human interaction.
Read paper ↗ 02 · ROUTINGFrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance
Chen, Zaharia & Zou · 2023 · Cost-aware model selection and cascades.
Read paper ↗ 03 · EVALUATIONAgentBench: Evaluating LLMs as Agents
Liu et al. · 2023, revised 2025 · Interactive agent benchmarking.
Read paper ↗RESEARCH IN PROGRESS
The standard is evidence
we can defend.
Methods are open for scrutiny. Results will appear only when the evidence exists.