What the factory has actually shipped.
Every project here is evidence, not a pitch. Each one exists to test an operating-model claim against something real, and each one's decisions are logged as it happened, not cleaned up afterward.
Hekton
The agentic engineering factory this whole site writes about, and the system it is built on: labs for unproven work, gremlins for the boring load-bearing tasks, and a human gate on anything that ships.
Read the thesis →The Gremlins
A roster of reusable, named agents scoped to one job each: reviewing, coaching, scaffolding, verifying, guarding. VeilGremlin is the sharpest public example so far: a local-first PII-masking layer whose mechanism validated end to end, then a human sign-off gate correctly caught that it could not deliver on its own governance promise.
See VeilGremlin on GitHub →Stealth Labs
The pattern behind a small internal platform for hosting confidential demo and prototype projects: remove the inbound network path instead of defending it, then reuse the result as one Terraform module per project. What stays private is the specifics, not the fact it exists; the generic pattern is public, with an interactive architecture walkthrough and a sanitized reference repo.
See how it works →Kriterion
A pre-registered test of whether five independent AI contexts, with structured challenge and preserved dissent, beat one strong AI context on evidence discipline. The aggregate comparison found no broad advantage for the committee. In one pre-registered framing test, the single-context recommendation changed while the five-context structure’s did not; the full margin-and-cost criterion remains unevaluated.
Read the findings →Looking for the workshops?
Coderturtle is the builder alter ego behind this site, running a different experiment: the hypothesis that harness-driven, agentic engineering needs new ways to learn, not just new things to learn. The projects on this page are evidence for an operating-model claim, built once and written up. The workshops at coderturtle.io are the opposite mode: hands-on, public, and run repeatedly to test how people actually learn to build with agents, in domains from a traditional programming language to loop engineering and context management.