What is Neural-Fractal Agentic AI?
The short version: OMEGA doesn't write one big prompt and hit send. It breaks every task into a tree of small, scoped cognitive units, runs each one with only the context it needs, and reassembles the answers.
Why "fractal"?
Every cognitive unit can spawn sub-units when the work gets too big for it. The structure that comes out looks like a tree: a root objective at the top, progressively narrower children underneath. Zoom in at any level and the pattern repeats — each node has its own observe-orient-decide-act loop with its own budget.
Why "agentic"?
Each unit is a first-class agent, not a prompt string. It has tools, a budget, a persona, and an approval tier. A research task can spawn a browser agent, a code task can spawn a shell agent, and the parent only sees the summary the child chose to return — not the raw 40-page intermediate.
Why it's cheaper
Token cost scales with prompt size and with the model that prompt is routed to. A single monolithic prompt for a complex task looks like this:
- Entire conversation history
- Full tool schemas
- All loaded documents
- Every system message
Because any one of those tokens might be the one that needs a frontier model's reasoning, the whole thing has to be routed to the most capable (and most expensive) model available.
The same task, fractally decomposed, is a dozen small prompts. Each one carries only what it actually uses, and each one can be routed to the smallest model that can handle it: a tiny planner, cheap per-source workers (often a local MLX model for free), a medium model for the final synthesis. On our reproducible 5-task benchmark, that works out to an 80–85% cost reduction vs. a monolithic frontier-model approach. (Full methodology and raw numbers live in benchmarks/token_efficiency/ in the repo — the benchmark is fully offline, anyone can reproduce it.)
An honest counter-point: fractal decomposition actually uses more tokens than monolithic on small tasks. Each level has its own system prompt overhead, and the synthesizer has to read worker summaries. The cost win is entirely from model routing, not from saving tokens. If every token had to go to the same model, fractal would lose.
Why it's smarter
Smaller prompts mean tighter attention. A well-scoped sub-task gets the model's full attention on the thing that matters, rather than splitting it across twenty pages of tangentially-relevant history. The accuracy gain comes for free once you're already doing the decomposition for cost reasons.
What this feels like to use
You type one message. OMEGA figures out whether it's a quick answer (one small model call, no decomposition) or a bigger task (spawn a tree, do the work, surface the result). You can watch the tree in real time if you want, or ignore it and just read the answer. Either way, the bill is a fraction of what it would be in a single-prompt system.