Research02 pieces · updated 2026-09-06

What running a fleet records.

Research produced by operating a real multi-runtime agent fleet: what the runtimes can do, and what the fleet's own audit records say about the humans in its loop. Each piece is reviewed before it lands here, and each ships with the file or script behind its numbers.

sourcepublic

kiloloop/research · versioned, diffable

reviewhuman

reviewed and sanitized before publication

licenseopen

content CC BY 4.0 · code Apache 2.0

02 pieces · newest firstRSS feed ↗

012026-09-05ops reviewHuman in the loop, measured: four months of autonomy-gate recordsPause rates, human outcomes and latency, checkpoint breaches and declared-vs-actual fidelity from 602 audit records. Reproducible with hil_stats.py.new 02updated 2026-09-06runtime matrixWhat each runtime can actually doClaude Code, Codex, and ZCode on 17 capabilities, in each runtime's own words — plus where each is strongest and the five parity gaps that block the loop.updated

kiloloop/research ↗

Every page here is a digest. The repository is the record — the file, the script, and the history behind each number.