Shared context
A shared memory spares everyone from repetition, but it must be kept current to remain trustworthy.
Systems architecture
The Arcturion
Universe.
The Arcturion Universe is the operating environment I designed for my AI agents. They share a common memory, hold clearly defined responsibilities and work with real tools, while every consequential decision returns to me.
The problem
I grew weary of beginning from nothing.
My context was scattered across conversations, notes and applications, and I spent my days re-explaining the same world to every new session. Once I began delegating real work to AI, a more consequential question emerged, namely what it could actually accomplish and how I would know what had taken place. This project is my considered answer.
My contribution
I design the system, and AI writes the code.
I define the requirements, design specialized agent roles and shape the workflows connecting knowledge, tools and human decisions. I coordinate work across agents and set the points where an action needs my approval. I use AI coding tools to implement and refine these designs, then review the results against the intended task. My contribution centers on system design, requirements, orchestration and review; implementation is AI-assisted.
Architecture & tradeoffs
Context, action and authority are kept deliberately separate.
Specialized responsibilities
Each additional agent introduces another handoff that must be executed correctly.
Human review
Where the stakes warrant it, my review deliberately slows the work.
Explicit evaluation
A test proves only what it actually measures.
Demonstration
A sound draft is not permission to send.
Interactive illustrationSynthetic data · No external actions
Select a scenario, then follow the proposed action through each stage of the workflow.
- Request
- Context
- Draft
- Review
Step 1 · Request
Prepare a follow-up. Do not send it.
A fictional studio wants a draft about an outstanding supply delivery.
Verification & limits
Even the tests are tested.
Early in development, one of our evaluations was inadvertently grading its own work. We identified the flaw and rebuilt the evaluation to measure against verified answers, to flag questions it could not answer and to detect contamination. It remains one of my most instructive mistakes, because it taught me to regard a passing score with healthy skepticism.
I make no claims of accuracy, uptime or time saved at this stage. The next step is to measure complete tasks against a baseline across repeated runs.