Agent-Augmented Practice
How does one architect run a multi-product studio? By designing the practice like a system.
Can one architect run a practice that ships software products, maintains a bilingual public site, runs a research lab, and still draws?
The honest answer: not by working harder. The hours don't exist. The only version of this that works is the one where the practice itself is designed: the way you'd design a building system, with load paths, redundancy, and inspection points.
This is a working note from inside that experiment. Since 2025, Fraktal has operated as a solo studio augmented by AI agents: coding agents that build and refactor, research agents that sweep and summarize, review loops that check public claims against reality. No staff, no outsourcing; one person and a system.
There are no invented numbers in this note. It describes patterns we actually run daily, and it is equally explicit about what breaks: context loss, taste, and the fact that every output still queues behind a single human reviewer.
The Loop: Frame the task in writing, delegate the legwork to agents, verify against reality, record the state. Repeat.
Theoretical Framework
Documentation as Source of Truth
One fact, one home. Company facts live in a single document; product state lives in per-product status files. Agents read these before acting, and update them in the same session when reality changes.
Claim-Check Discipline
No sentence goes public that contradicts the code or the live product. Before publishing, agents grep the site for claims (prices, statuses, feature promises) and check each against the source of truth. This lab page went through that same pass.
Task Orchestration
Big work is written down as a plan before any code moves. Phases get delegated to parallel agents; independent tasks run concurrently; everything converges on one review.
State Files as Handoff
Status files carry the practice between machines and between sessions. An agent on a different computer, or a different week, reads the state and continues. Memory lives in the repo, not in anyone's head.
Research Process
Frame
The task is written down first (context, constraints, definition of done) before any agent touches code
Delegate
Independent phases go to parallel agents; the architect stays on decisions, not keystrokes
Verify
Builds, tests, and claim-check sweeps run before anything is called finished
Record
Docs and status files are updated in the same session; the next session starts from written state
Research Phases
Ad-hoc Prompting
Where everyone starts: asking an assistant for snippets and drafts. Useful, unrepeatable, and it taught us what breaks without structure.
Rules in the Repo
Writing the studio's constraints into the codebase itself: brand rules, claim-check rules, per-product instructions. Agents stopped needing to be told twice.
Orchestrated Runs
Written plans, parallel delegated phases, one consolidated review. The current mode of operation; this very page was produced this way.
Practice as Product
Packaging what we learned as a service for other architecture offices: their visual language, their data, their agents. Sold honestly as a pilot.
Key Metrics
Key Thinkers
Cedric Price
Price treated buildings as systems that respond and reconfigure; the Fun Palace was a machine for change. We apply the same stance to the practice itself: not a fixed office, a reconfigurable system.
Christopher Alexander
A Pattern Language showed that good environments come from named, reusable patterns. Our agent skills and repo rules are exactly that: patterns written down so the system can apply them without us.
Archigram
Archigram drew practices, not just buildings: walking cities, plug-in living. A one-person studio running on agents is closer to their drawings than to a conventional office.
Buckminster Fuller
Ephemeralization: doing more and more with less and less. Fuller meant material; we find it applies to organizational mass too. The lightest practice that still ships.
Where This Lives in Our Products
AI Systems: this practice, sold
PilotEverything in this note is the substance of our AI Systems service: we set up the same documentation discipline, claim-check loops, and agent workflows inside other architecture offices. Offered as a pilot, priced as a pilot.
AI Systems →The products themselves
Daily useArchly, SpaceCraft, and Falcon AI are built and maintained through this workflow. The practice is not a demo of the method; the method is how the practice exists at all.
Products →Key Findings
Documentation stops being bureaucracy when agents read it. A status file nobody reads is dead weight; a status file that configures your workforce every morning is infrastructure.
Docs became load-bearingState files make the practice portable. Any machine, any session: the agent reads the written state and continues. 'Where was I?' became a solved problem.
State survives machinesClaim-checking has to be systematic, not moral. Good intentions don't catch a stale price on a product page at midnight. A grep does.
Honesty as a pipelineParallelism moves the bottleneck, it doesn't remove it. Five agents working means five outputs waiting for one reviewer. The practice scales exactly as far as the founder's attention.
Review is the ceilingHonest Limitations
Context loss is real. Agents forget between sessions; the written state catches most of it, but nuance leaks. Some decisions get re-argued because nobody remembers they were settled.
The review bottleneck is the founder. Every public sentence, every design call queues behind one person. On a bad week, the queue wins.
Agents are confidently wrong about taste. They can verify a build, not whether a facade proportion sings. Aesthetic judgment has not been delegated, and maybe can't be.
This note describes months of practice, not years. Whether the system survives real scale (more products, actual clients in the loop) is exactly what we're finding out.
Conclusion
A solo practice with agents is not a smaller version of an office; it's a different organism. Documentation is its skeleton, claim-checking its immune system, and the founder's attention its hard metabolic limit. It works today at this scale, honestly stated. What we don't yet know is how far it stretches, which is why this is a working note, not a conclusion.
Limitations
- Single review gate
- Months of evidence, not years
Future Directions
- Measuring where founder attention actually goes
- Piloting the system inside other offices