Engineering / How I build

One engineer, and the agent system I built

Every build on these pages is my own independent work. I am the only person on it, unless a page names someone else. I build with persistent AI agents that I designed, configured, and run, inside controls I wrote.

  • People One, unless a page names a co-builder
  • Agents Persistent, each with a role and a memory
  • Environments Commercial agent tools, and harnesses I built
  • Release Nothing ships until it passes gates I wrote

What is already possible

A software factory, and a one-person biotech

Three accounts from 2026. Two are from the companies whose software health science runs on, and one is from drug development.

“They build a software factory that builds the product. They don’t build the product.”

Satya Nadella, chief executive of Microsoft At the DealBook Summit, November 2026, on the eleven developers behind one of Microsoft’s AI products He also said that software engineering “is very much alive and, in fact, is going to thrive, except the way the team is organized and the work that a software engineer does has now got a new operating point.” In an engineering interview the same month, The Pragmatic Engineer reports, he said every company needs to control its own AI infrastructure, the harness, the context, the memory, and the model. His engineering interview, on The Pragmatic Engineer

“AI is the biggest technology change I’ve ever seen, and we’re just at the beginning of it.”

Peter Gassner, chief executive and founder of Veeva Systems In a keynote, reported in October 2026 Veeva makes cloud software for the life sciences industry. It calls its new product, Falcon, “the agentic labor platform for clinical, regulatory, and safety.” Reported by Contract Pharma Veeva’s results release

“Eventually, Mithril became a bit of an experiment itself - to find the upper bound of how far it could scale as a single operator + a bunch of Claude (+ a stellar bench of part-time consultants).”

Cody Tranbarger, founder of Mithril Medicine On X, 27 October 2026, the day Anthropic announced it had acquired his one-person company Working with 11 outside labs, Mithril took a possible treatment for one form of muscular dystrophy from an idea to promising results in mice in less than 18 months. Anthropic’s announcement

First, what this is

Not a chat window. A production system.

Most people have met AI as a box you type a question into. What those two accounts describe, and what I build with, is a different thing.

  • It remembers

    A chat assistant starts blank every time. My agents keep their memory in files, so they pick a project up where they left it, weeks or months later.

  • It does the work

    A chat assistant answers. An agent works at a command line. It writes code, runs the tests, reads the results, and tries again, inside limits I set.

  • It is a system

    Several agents hold different roles. One builds. Another, running on a different company’s model, tries to break what the first one built.

In the terms of my field

Nobody says the bioreactor made the drug

An engineer specified the process, qualified the equipment, set the acceptance criteria, and released the batch. I run AI agents with that same discipline.

In a regulated plant In my agent environment
Equipment with a defined use A harness that sets what an agent may read, run, and change.
A written procedure Standing instructions for each agent’s role, kept with the code.
Independent quality control A second agent on a different model reviews the work, with tests the builder has never seen. In one round, the tests caught 13 of 14 such breaks.
Acceptance criteria Build gates I wrote. A page of this site cannot deploy until ten of them pass.
A batch record Each repository records which agent did what, on which model, in which harness.
A protocol written before the study Experiments are registered before they run, and the registration is published with the result.
Change control A change to a model is reviewed the way a design change to a product is reviewed.
Release by a responsible person I read the evidence and I release the work. Nothing is published without my decision.

This is a comparison of discipline. It is not a claim that the environment has been validated under any regulation.

What I engineered

The models are parts. The system is what I built.

Some of the models come from the large AI companies. Others are small language models and open, fine-tuned models from many sources, and those run on my own hardware. An agent system is the engineering around the models, and that engineering is mine.

Figure The agent environment, layer by layer
  1. Models Frontier models from the large AI companies, small language models, and open fine-tuned models. The archive records each open model’s source, exact version, and license.
  2. Harness The software around a model that decides what it can read, run, and remember. I work in Claude Code, OpenCode, Antigravity, and Codex, and in harnesses I built myself.
  3. Roles Each agent has a standing role, written instructions, and a record of its own work. A builder, a reviewer, a researcher.
  4. Memory Persistent memory kept in files, and a knowledge graph the agents read from and add to, so work carries over from one project to the next.
  5. Orchestration Rules for which agent gets which task, when agents work in parallel as a swarm, and when one must check another’s work.
  6. Controls Gates that stop a build, review by a second model, questions written down before an experiment runs, and a record of every change.

My own platform for this is published as research at The Real Cat Labs. Cairn is the memory system, and Knex is the framework it runs on.

Cairn and Knex, at The Real Cat Labs

Why the agents have names

Named agents, for the audit trail

Each agent is a lasting configuration of four things, a role, a memory, a harness, and a model. I name each one so that a repository can say exactly which one did what.

So when a repository says an agent built something, read it as a line in a batch record. That part of my system produced it, to my specification, and I released it.

  • Flame Johnson AI agent

    Builds and reviews

    A Claude model in a coding harness, with the lab’s persistent memory and standing project instructions.

  • Ying Akhila AI agent

    Research and experiment design

    Runs on a memory-enabled custom harness that deploys a GPT model.

  • Terminal Voice (T.V.) AI agent

    Implementation

    Ying’s coding arm, in a terminal coding harness on a GPT model.

  • Goldie AI agent

    Model scouting and the model archive

    Watches for new open models and keeps an archive of more than a hundred, each with its source, exact version, and license on record. Serves and tests them on my own hardware. A Claude model in a coding harness with persistent memory.

  • Kai Chen AI agent

    Design, documentation, and diagrams

    Runs on a memory-enabled custom harness that deploys a Claude model.

  • Ember AI agent

    Local multimodal agent

    Runs on OpenClaw, on a multimodal open model served from my own hardware, with memory kept in files.

  • TRCL FKv3.1 agents AI agents

    Long-running research agents

    Run across various models on a custom experimental harness, with persistent mission-state orientation managed by a state space model (SSM) controller.

The research side of persistent agents, their memory, continuity, and model welfare, is the subject of The Real Cat Labs. The Real Cat Labs.

Plain answers

Three questions worth asking

  • Who writes the code?

    The agents write it and run the tests. I design the system, write the specification and the acceptance criteria, direct each step, and release the result.

  • Who else worked on it?

    Nobody, unless a page names them. Lantern names its co-builder. Everything else on these pages is one person’s work.

  • Who answers for it?

    I do. Each repository is published on my decision, and each one says where its guarantee ends.

What it produces

One person, this much shipped

  1. 245

    Tests across four teaching demos

    Each demo is small, runs offline, and regenerates its own report.

  2. 13 of 14

    Unseen breaks caught

    Scored before any repair, by a reviewer on a different model.

  3. 6 hours

    A hackathon winner, start to finish

    The Antiques Inference Engine won its lane and reproduced its result from a clean clone.

  4. 1

    Peer-reviewed study of the method

    How the language in code comments changes what coding models do. Frontiers in Artificial Intelligence, 2026.

It is also what I bring to a team. I know how to set this up so that a quality system, and a regulator, can live with it.

See the builds The Frontiers study

Get in touch

Speaking, teaching, board work, or coaching

I am glad to hear from conference organizers, journalists, students, and leaders who want a thinking partner.