Software architect in Santiago, Chile. Thirteen years on banking, telecom and public-sector systems. On my own time I run AI coding agents on real code and measure what happens: what they cost, how much of their code survives, where they get out of a sandbox. The notes are at carlostapia.cl.
Tools
- seisin: gives each agent its own folders and keys, enforced by the OS. MIT, on npm.
- tartaros: a local sandbox for coding agents, in Rust. Apache-2.0.
Measurements
- sandbox-conformance: runs the tests other sandboxes wrote for themselves against every sandbox, with the results by date.
- denial-detours: what an agent does when it is denied a write to another role's file. Exploratory, every trial kept.