I run a fleet of AI coding agents on real client and product work, and I publish the log.
Most agent failures are not reasoning failures, they are operating failures: the agent starts blind, grades its own work, and writes nothing down. I build the small pieces of plumbing that fix that, and I run them in production before I publish them.
One 30 day window ending 21 August 2026, counted from my own append only ledger. Not a benchmark, not a demo run. The whole table is recut only when the next full window closes, never partially.
| Work lanes opened | 313 |
| Closed as done | 283 (90.4%) |
| Ledger entries written | 725 |
| Written by | Claude 560, Codex 165 |
Where those lanes went: client and product 114, infrastructure and fleet operations 111, research 33, creative 20, publishing 19, other 16.
lanes is the coordination layer the fleet runs on. Take a lease before you touch files, produce proof before you close, leave an append only trail the next session can replay. Python standard library, zero dependencies.
royceos is the operating model underneath it: boot from a fixed source of truth, let one model build and a different one verify, require runtime evidence, hand off durably. Plain Markdown and shell, vendor agnostic.
roycecode is native Rust whole codebase analysis for agents. It builds a semantic graph and emits machine readable artifacts, so a reviewer reads evidence instead of the entire repository.
- X: @vojtechroyce
- LinkedIn: vojtechroyce
There is no token, no coin, no paid DM, no investment offer, no giveaway, and no "assistant" who messages you on my behalf. I will never contact you first about money. Official links live only here, on this profile. Anything else claiming to be me is not me.