The Loop: how AR Labs works

The Loop is how AR Labs works: a cycle of five stages, drawn above, in which we learn something, experiment with it, narrow the experiments into one working system, apply that system to a real problem, and tell the story of what worked and why. The story feeds the next round of learning. Each section below says what we do at one stage and points to work on this site that came out of it.

At the centre of the drawing is "AR LABS: the visible surface". This site is that surface: most of its pages are the output of one stage or another, so the cycle can be read from them.

01 Learn: education, new technologies, emerging patterns

In the drawing's legend, education feeds experimentation. We study before we build, and we study in a form that can be checked later. The research library holds dated reports we carried out before engineering decisions, such as which framework to build a voice agent on and how to keep long-running AI agents effective as their context fills. Most of these reports were run with ResearchBot, which scores every source with the CRAAP test and needs three independent sources for each major claim, or with multi-agent research, which splits one question across parallel researcher agents.

Structured study runs beside the research. Richard M. Thompson, who founded AR Labs, holds Anthropic's Claude Certified Architect credential, and we built a small study tracker while he prepared for it. The habit goes back to a one-year self-directed syllabus in software and AI. Learning also happens inside a build: when a build shows something we cannot explain, building pauses until we understand it, so that AI tools do not produce code faster than the people responsible for it can follow. Code Tutor and Comprehend are the skills we use for that.

02 Experiment: platforms, languages, methodologies, technologies

In the legend, experimentation narrows to production. An experiment tries a platform, a language or a method at small scale, to find out what it is good for. VoxPlan, an open-source Android app for goal planning, and two voice-first Android apps, Vox Manifestor and Coherence, were experiments in voice interfaces on the phone. Bpai Dûay Gan (Go Together) was a hackathon app. The Code Health Skill prototype tried a multi-pass review method for agent-written code.

Methods are experimented on as well. The coding pipeline, the route every code change takes from spec to a person's merge, runs on our own development first, and each of its rules was added when work showed it was needed.

Several of these projects are paused or finished. What carries forward is the pattern that worked: Coherence's spoken check-in, in which one role talks and a separate role checks whether the required information has been gathered, became the model for the redesigned conversation in Salesbot.

03 Produce: narrow focus, build the demo, ship it

In the legend, production finds application. At this stage we pick one thing, build it until it works end to end, and put it where people can use it. We build working demonstrations because a working system shows more than a description.

Salesbot is the clearest example. It is a browser voice agent built on Pipecat that runs a short sales discovery conversation, reads back what it heard and closes with a summary, and anyone can talk to it. Break It Down, an open-source Claude Code command that turns a long AI response into a step-by-step conversation, shipped the same way.

Much of what we produce is tooling for our own work. Taskgraph explains where a branch of the task graph stands, and VOX Torch and Vox Canvas draw that graph in the terminal and the browser. We run our own work on these tools, so each one meets real work before anyone else's.

04 Apply: find businesses and organisations it helps

In the legend, application becomes story. Applying a method means taking it to a business with a problem we did not choose. The music school case study is the fullest example: for a regional in-home music school, we mapped how an enquiry becomes a teacher match, a trial lesson and a weekly slot, and designed a workflow automation system that keeps the owner in charge of every pairing decision. It is a design; nothing in it has been built. The discovery method we used to map that school's real workflows is written up on its own page.

The guides carry the same method to other small service businesses, from tutoring companies to trades firms, and recommend the simplest change first, before any software or AI. Applied work also tests our own habits: a system shaped around one person's way of working may not suit anyone else, and work for a business shows early which of our assumptions were personal.

05 Share: tell the story of what worked and why

In the legend, the story feeds the next cycle. We write about what we build. The Salesbot project page explains how that voice agent works and how it was built, and the blog carries the essay series that begins with The Acceleration Paradox. The building in public skill turns daily engineering work into articles. The pages in this lab, from engineering practices to research reports and case studies, are the same story told in more detail.

Explaining a piece of work shows where our understanding of it is thin, and that is where the next round of learning starts.

Self-developing, self-funding, self-educating

The foot of the drawing names three things we intend the cycle to do. It is self-developing: what the Produce stage ships for our own use, such as the coding pipeline and the task-graph tools, does the work of the next cycle, and a mistake an agent repeats becomes a check in its environment, as harness engineering describes. It is meant to be self-funding: the Apply stage is where work for businesses is meant to pay for the learning and experiments that come before it. And it is self-educating: we choose work so that building it teaches the skills the next piece of work will need.

Together they describe a small practice that keeps improving its own tools, which is the point of our mission to put this kind of capability in the hands of individuals and small teams.