The NFA loop, explained.

In April I published a primer on Neural-Fractal Agentic AI™ (NFA), the engine underneath every request OMEGA handles. That post made the argument: many small, scoped units of work beat one giant prompt on cost and on accuracy. This post is the companion piece I promised. It follows a single request through the loop it actually travels, from the sentence you type to the file that lands in your folder. Six stages: goal, think, route, act, deliver, learn.

You never have to know any of this to use OMEGA. The loop runs whether you watch it or not. But I have noticed that people who understand what happens at each stage start asking for bigger things, because they stop treating the product like a chat window and start treating it like a colleague with a process.

Stage one: a goal in plain words

Every request begins as ordinary language. "Turn Tuesday's call notes into a proposal deck in our brand and remind me to send it Friday morning." No special syntax, no prompt tricks.

Most AI tools still quietly reward prompt craft. The better you phrase things, the better the output, which means the tool is grading you. I think that is backwards. OMEGA's first job is to work out what done looks like for your sentence: which call you meant, which brand assets apply, what a proposal deck contains, and what Friday morning means on your calendar. If the goal is genuinely ambiguous it asks one clarifying question instead of guessing. Your sentence is treated as a destination, not as text to be continued.

Stage two: think, which means recall before planning

Before OMEGA plans a single step, it checks what it already knows. Memory here is four layers deep, from the immediate conversation up to a persistent knowledge graph we call the Brain. Yours holds your projects, your people, your preferences, and the decisions you have already made, so "our brand" resolves to your actual Brand Kit and "Tuesday's call" resolves to a specific set of notes rather than a guess.

Then it plans. The goal gets broken into small steps, and stubborn steps get broken again, which is the fractal part of the name. Each step ends up scoped tightly enough that OMEGA can check whether it worked. A plan whose steps can each be verified is the difference between an assistant that finishes jobs and one that produces confident partial work.

ChatGPT's memory feature and Claude's Projects both do real recall, and both are genuinely useful. The difference is what the recall feeds. In a chat product, memory shapes the next reply. Here, it shapes a plan that the rest of the loop will execute. There is also a separate switch for slow, deep cognition, Quantum Neural-Fractal Agentic AI™ (QNFA), which I covered in its own post. Everything below is the standard loop, with that switch off.

Stage three: route, or matching each step to a mind

This is the stage I most want people to understand, because it is where the money is. Once a plan exists, every step gets routed across four tiers of intelligence:

• A local tier that costs nothing per token, because it runs on your own Mac through Omega-MLX on Apple Silicon.
• A fast, inexpensive cloud tier for simple steps that still want a bigger model.
• A balanced tier for ordinary reasoning and tool use.
• A deep tier, reserved for the few steps that are genuinely hard.

The industry has already conceded the premise. OpenAI ships its GPT-5.6 line as three models at three prices. Sol, the flagship, runs $5 in and $30 out per million tokens. Terra, the balanced model, is $2.50 and $15. Luna, the fast one, is $1 and $6. Anthropic prices Claude Fable 5 at $10 and $50. Nobody prices intelligence flat, because intelligence does not cost flat. The open question is who picks the model for each piece of your work. In most products you pick once, per chat, by hand. In OMEGA the router picks per step: the right agent, the right model, the right cost.

A word for the other side. One big model for everything is a defensible design. It is simpler to build and simpler to explain, and today's flagships are remarkable. Give Fable 5 its million-token context window and it will hold more of a project in view at once than anything before it. If your day is one hard problem and cost is no concern, a flagship-only setup will treat you well.

Routing wins on mixed work, which is what real work is. A proposal deck is twenty steps, and maybe two of them are hard. Send all twenty to a flagship and you pay deep-reasoning prices to reformat bullet points. Send all twenty to a cheap model and the two hard steps come back shallow. Routing puts the expensive model exactly where it earns its price. Quality improves for a second reason the primer covered: a small, focused prompt gets a model's full attention, while a giant one splits attention across everything you ever pasted in.

Two mechanical notes that matter in practice. You can bring your own API keys, and we add no markup on tokens; you can also route through a Claude or Codex subscription you already pay for, and on Apple Silicon a surprising share of steps costs nothing at all because it runs locally. And routing is provider-portable: you can swap models or providers mid-conversation without losing context, because your memory belongs to OMEGA rather than to whichever vendor happens to be answering this minute. Budgets cap what any run can spend, so an ambitious plan never surprises you on the bill.

Stage four: act

A plan is worth nothing until something touches the world. OMEGA ships with more than 200 specialized agents and over 5,300 bundled skills, so acting can mean reading and writing the files on your Mac, calling the services you have connected, filling a spreadsheet, or operating real applications through computer use, where every app requires your explicit per-app approval before OMEGA may drive it. Voice is processed on-device, and when a task needs a human on the other end of a line, OMEGA can place an actual phone call.

Recurring work becomes an Automation. You describe the job in plain English, OMEGA builds the triggers and schedules, and the result runs on an always-on runner, so it keeps working while your Mac is asleep. If you already live in n8n, your existing node graphs import directly.

The competition at this stage is serious. Zapier connects more than 7,000 apps, the deepest catalog anyone has. ChatGPT Work arrived this month with over 1,400 connectable apps and scheduled tasks. Claude Cowork reads and edits files in folders you choose and now runs background agents from mobile and web. These are strong products from teams I respect. What I would push back on is the shape: in those products, acting is the product. Here it is one stage of a loop, so every action inherits the recall from stage two and the cost controls from stage three. It also owes stage five a finished result.

Stage five: deliver

My standard for this stage is blunt: did a finished thing land somewhere you can use it. An answer buried in a chat transcript is not a deliverable. OMEGA exports across 47 formats, slides, documents, spreadsheets, entire books, with your Brand Kit applied, so output arrives in your fonts and colors instead of default-template gray.

Bigger deliverables get their own rooms. Design Studio is a full AI-guided UI/UX studio: an infinite canvas with frames and design systems, plus a copilot that proposes edits you review and can undo. Project Canvas takes an idea through a plan you can see as nodes, into real code with a live preview, readable diffs, a terminal, and a deploy step, and it accepts existing work, whether that arrives as files, a local folder, a git repository, or a Figma link. The two bridge, so a design can walk straight into a build.

Claude's Artifacts and ChatGPT's Canvas editor deserve credit for pushing chat products toward real output, and for a single document they are pleasant. The gap shows when the deliverable is plural: a deck and the spreadsheet behind it, both on brand, both actually saved where your work lives.

Stage six: learn

The last stage is the quiet one. When a job finishes, OMEGA writes back what happened: which plan shape worked, and what you corrected before you were satisfied. The Brain graph grows a little. Goals let you leave standing objectives that persist across sessions, so progress on long work continues without you re-explaining it. And overnight there is Dreams, a self-improvement pass that reviews recent work while you sleep, so the next first draft starts closer to your taste.

This is the stage that compounds. A chat product is the same product on day 200 as on day one. OMEGA on day 200 has been corrected by you for months, and it shows in how rarely you have to repeat yourself.

One request, around the loop

Put it together with the example from the top. You type: "Turn Tuesday's call notes into a proposal deck in our brand and remind me to send it Friday morning."

Goal resolves the ambiguity: which call, and which template counts as your brand. Think recalls the notes, plus the last proposal you approved as a structural reference. Route sends outlining and reformatting to models that cost pennies or nothing, and reserves a deep model for the one section that has to persuade. Act reads the files and builds the slides, then creates a small Automation for the Friday reminder. Deliver puts a finished deck in your folder, in your fonts. Learn notes which sections you kept and which one you rewrote, so the next proposal starts from a better first draft.

Nothing in that run required a prompt trick or a settings page. That is the job NFA exists to do, and it is why the product feels different in month two than in week one. OMEGA runs natively on macOS 14 and newer on Apple Silicon; the download page has the details, and pricing covers what it costs to run.