Arcanada month three. 823 core tasks and a Mac I can turn off

Arcanada is now in its fourth month, which makes it time to look back at the third. This report covers June 15 to July 18, 2026. A lot happened, with even more still ahead, so I will stick to the main results.

Arcanada completed 823 core tasks in month three.

That is 142% higher than the result for month two.

The 823 core tasks include 492 Datarim task IDs, 164 separate Cubrim hypotheses, and 167 Angry Robot hypotheses. The 3,711 routine tasks consist of 3,671 Angry Robot backtest runs and 40 independent full Cubrim benchmark runs.

Month-three costs

Over three months, I have put at least $2,240 into Arcanada.

Control is gradually moving to the agents

Control of Arcanada is gradually moving to the agents. I still know how the system is put together at a high level, but without help from the assistant I cannot always say how a particular function works or how individual components interact. Many tasks run autonomously in the background. Orchestrators keep launching agents, and the agents work around the clock.

At one point, the backlog passed 2,000 tasks by my estimate. Some had already become outdated and needed review. I handed the accumulated backlog to the orchestrators to work through. In my experience, the models have slowed down, and one Claude subscription was no longer enough for this workload. I added Codex and Cursor. I use Cursor's $20 monthly plan in Auto mode for simple programming modules, quick DevOps tasks, and consilium work.

When agents build interfaces, I still have to go back and fill in gaps or correct imprecise design decisions.

The Mac can now be turned off

All Arcanada development now takes place on development servers in data centers. No task runs on the Mac anymore. I sometimes leave the Mac on just for monitoring because it is convenient to watch the work there. The tasks themselves no longer depend on it.

If I turn off the Mac, the work carries on. The agents and orchestrators continue running tasks on the servers. For me, this is a major achievement of Arcanada's third month.

The ecosystem and commercial projects

The number of Arcanada projects increased severalfold this month. Some of the new projects are small open-source libraries reused elsewhere in the ecosystem. Others remain closed commercial products.

I am still developing Verdicus, Transcribator, and Model Connector, including their user interfaces. Before I can release them, they need marketplace approval, the legal issues need to be resolved, payment accounts need to be connected, and the publishing keys need authorization.

With past projects, problems around publishing and app authorization forced me to change infrastructure, domains, and sometimes even the company name.

These procedures are not always under my control. I work on Arcanada alone, so I move the commercial work, interfaces, and releases forward gradually.

Cubrim and Cubrim-2

I spent less time on agent systems and the central control panel this month because I focused on one of my older ideas. Most of month three went into building a working Cubrim archiver. By my weighted assessment, it reached first place in the world ranking. I am proud of that result.

Building the archiver took more than two weeks. But the archiver is not the end goal of this research. The original plan was to take any archiver that was good enough, pack the data, and then reduce that packed volume even further.

That second part needs a global addresser. Cubrim-2 focuses on that problem. The archiver already works on its own. The global addresser is the next research stage. I am continuing the research, testing hypotheses, and looking for improvements.

Scrutator, memory, and Wiki Arcana

This month I got Scrutator working. It manages the knowledge bases for the agents and the ecosystem. It brings long-term memory, vector storage and search, knowledge graphs, retrieval of the needed information, and result checking into one system. I expect this approach eventually to save agents up to 90–95% of their tokens by giving them only the information they need for each task.

Wiki Arcana is next. It is a larger public system patterned after the Jedi Archive. Knowledge is organized into roughly eight levels, from individual projects and modules to project spaces and broader areas. Every piece of knowledge and every graph gets pointers describing its location and ownership.

Wiki Arcana brings long-term memory, Scrutator, and knowledge graphs into one place. Access will depend on role. I want to standardize this system and open it to the public.

The future Arcanada agent and Muneral

Knowledge graphs and memory standards will be the basis for Arcanada's own agent and its task tracker. The agent needs to be able to form knowledge, find it, read it, and use it while working on tasks.

I still use Datarim's file-based task system. In month four, I plan to launch my own multilevel Arcanada agent for the first time. After that, I will launch Muneral, the agent task tracker. The first version will be read-only. Over time, Muneral should make the ecosystem's work more structured and easier to see, while helping me regain control.

Datarim

Everyday work kept generating follow-up tasks in Datarim, covering improvements to skills, agents, approaches, rules, and the framework itself. That created a large development backlog.

Most of month three went into those accumulated Datarim improvements. I cleared the main backlog, with several dozen tasks still left. I use Datarim every day. It already has orchestration and level-five autonomy, so it should need far less of my active attention from now on.

I am gradually moving my focus from Datarim development to Muneral and a higher level of managing agent work.

Publisher

Publisher is an open-source framework that automates publishing prepared content through browsers. Social networks fight bots, so publishing has to happen in a browser. Agents had a hard time with Playwright. They forgot the publication language or missed the need for a translation, chose an image instead of a video, and got confused by comments.

Publisher covers Telegram, X, Facebook, LinkedIn, VK, YouTube, and the Arcanada website. For each publication, it prepares text, images, audio, and video. I am integrating it into the Arcanada control panel so it can distribute prepared material to the right platforms with one action.

Preparing, editing, translating, formatting, narrating, and producing video all take time. I am gradually automating that process.

Plans for month four

1. Launch the finished projects in production, publish them in the Apple and Google marketplaces, get the first registrations, and possibly receive the first payments.
2. Bring the Arcanada control panel to the point where outside users can view my projects. Creating their own spaces will remain the next stage because that mode needs more testing.
3. Build a fully automated process that turns my drafts into articles and other text, prepares audio and video, and publishes the material to social networks and video platforms.
4. Launch Arcanada's own AI agent.
5. Launch Muneral, the agent task tracker. The first version will run in read-only mode.
6. Launch Wiki Arcana as a shared space and framework for working with knowledge, inspired by the idea of the Jedi Archive.
7. Hold offline negotiations and presentations for Arcanada and its projects.

The commercial side of a product

My conclusion from this month is not about agent systems. Sales, distribution, security, legal work, payment systems, monetization, and profit all need to be designed from the start of product development.

Doing this before development, or alongside product creation, increases the chance that sales can start with the MVP. Otherwise, a technically finished product may sit for several more months waiting for distribution channels, legal arrangements, payments, and a way to make money.