Eryk Orłowski · hretheumhretheum.com/questions/built-with-agents
Question P6 of 8
All questions6What have they built with agents on their own?
Comes up with, designs and builds their own tools with AI agents, alone. Kelter, an engine for running agents, has 291 tests and 32 recorded architecture decisions, and does the same work on a local 0.5B model and on a frontier model after changing 1 line of configuration. Gatefold, a design process with AI in which a person makes every decision, was built in 131 commits between 4 and 13 October 2026, and 12 patterns accepted in it became a course on this site's subdomains. The engine of this site gives only facts from the claim package; in a measurement on a real model, the chat gave no fact from outside the package.source sources: om-05, ke-01, gf-01, fo-01, ke-04, ke-03, gf-04, gf-05, fo-02, fo-04
Evidence
Kelter has 291 tests and 32 recorded architectural decisions.”
Measured on the real model, the chat gave no facts from outside the package, fell back in 3.3% of answers and took 9.5 s for the slowest 5% of answers.”
It depends on what hands-on means to you. […]It depends on what hands-on means to you. For me today it means building. I build my own tools with AI agents: Kelter, an engine for running agents; Gromatik, a platform for maintaining design systems; and a set of twelve tools that measures brand consistency for a client with many brands. I came up with them, designed them and built them myself. So I'm not someone who stopped making things and now only draws org charts. The kind of making has changed, and I'd argue it's the kind a team going into AI needs most.”
12 patterns accepted in Gatefold became the course on this site's job-posting subdomains, in production since 10 October 2026.”
On this site every fact comes from a profile package of claims with an identifier, a word-for-word quote and a status; only confirmed claims go in.”
Gromatik was built in 256 commits between 5 August and 13 September 2026, with about 1000 passing tests.”
I build tools with AI agents myself: Kelter, Gromatik and Exocortex.”
I know CI/CD principles from practice: in my own projects I run pipelines across dozens of repositories: tests and image builds, live contract checks, releases that can be rolled back, automated dependency updates and fork syncing.”
The toolchain is a monorepo of 12 command-line tools with 78 test files, built in June 2026.”
I came up with, designed and built the toolchain myself, working with AI agents.”
I came up with, designed and built Kelter myself, with AI agents; it is my intellectual property.”
In Kelter the kernel is the runtime and the language model is only a policy.”
The same job runs on a 0.5 B local model and on a frontier model after a change of 1 line of configuration.”
Kelter will be used in Exocortex R&D and eventually opened to the public.”
I designed and built Gromatik and blind.ds myself, as a private project.”
Generating component code is one model call per component, with no agent loop and no code execution on the server.”
Every run records its settings and a fingerprint of its input, so you can trace where a result came from.”
Exocortex pre-synthesises threads from different sources around a client, a project, a professional domain or a field of knowledge.”
Several dozen documents a day go through Exocortex in total.”
The lab is a separate, public project that uses only the Exocortex engine and is cut off from confidential and private data.”
The knowledge graph runs on PostgreSQL with the Apache AGE extension.”
I designed and built Exocortex myself.”
The hypothesis card is frozen and its checksum published before anything is calculated.”
No file reaches the public repository without 5 automated checks of the publication gate.”
The publication gate is tested every night on files with deliberately planted leaks.”
The lab's isolation from the private instance is checked every night by a test from inside the lab.”
In the first full week 198 papers went through the opportunity radar.”
I am the project owner and run the execution through agents.”
I designed and built Halvern, a fork of the open-source Meetily, working with AI agents.”
Halvern records, transcribes and summarises meetings entirely on the user's computer, with no account and no upload.”
In Halvern onboarding asks which language meetings are in, and the answer picks the transcription engine.”
Halvern checks for updates once at start-up, not on a timer, so it does not reveal when the user is in meetings.”
179 of 310 commits of Halvern were co-written with Claude.”
Halvern is a signed and notarised pre-release for Macs with Apple Silicon, with a public repository under the MIT licence.”
I came up with, designed and built Gatefold myself, working with AI agents.”
Gatefold was built in 131 commits on the main branch between 4 and 13 October 2026, with 32 recorded architectural decisions and about 500 tests.”
Designing 3 motion sheets with an agent cost 0.80 USD, passing 8 ready scripts through the agent cost 3.34 USD, and a plain program now does that step for 0 USD.”
I came up with, designed and built the engine behind this site myself, working with AI agents.”
For a job posting the engine matches each requirement to up to 3 claims and shows requirements without a match openly as gaps.”
216 of 235 commits of the site rebuild since 1 October 2026 were written with Claude.”
Before a merge, a gate in CI compares 60 views of the live site to prove that design concepts do not change it.”
Cognitive Mirror is an AI session that diagnoses how an adult learns and ends with one change to make.”
In Cognitive Mirror the model only collects observations, and a rule in code works out the score and confidence level.”
Splitting the analysis into a separate call per dimension raised the small model's agreement from 20 to 25 of 26 points.”
In the paraphrase test the production model gave the same result in 39 of 40 cases.”
The Cognitive Mirror interface is designed in stages closed by my decision, in 4 languages, and checked by 7473 automated tests.”
I see my own projects as my R&D rather than a second job competing for time. […]I see my own projects as my R&D rather than a second job competing for time. They're where I learn things before I need them at work. Building an agent engine and a design-system platform taught me how AI systems fail in practice, and that's exactly what I bring to a team adopting AI. The habit of setting a success criterion before measuring anything comes from my own lab and now shapes how my team decides whether an AI workflow is worth keeping. So the honest answer is that those projects make me better at this job.”
CI/CD pipelines are an everyday thing for me, because almost every project of my own has several. […]CI/CD pipelines are an everyday thing for me, because almost every project of my own has several. I run them across dozens of repositories on GitHub Actions, and the approach differs by project. In the engine behind this site every change goes through tests, a visual-parity check and end-to-end tests, then a container image is built, and the deploy remembers the previous version, so a rollback is a single command. In the demo apps for web, iOS and Android each platform has its own pipeline and its own check guarding the main branch. In projects with AI agents I batch changes into queues tested on separate branches, and the main branch moves only after a green result. On top of that come automated dependency updates, releases with installers and fork syncing.”
Limit
Cognitive Mirror has had no real users, and its instrument is not empirically validated.”
E-mail: [email protected] (kt-01) · LinkedIn: linkedin.com/in/eofek (kt-02)