🧩 Fundamentals
The "new way" and the vocabulary for understanding it: what it means to orchestrate multiple LLMs, the two questions that define any orchestration, and the spectrum from "you decide everything by hand" to "a trained model does the orchestration for you."
Every orchestration — manual or automatic, Fugu or Fusion — is an answer to these two questions. Remember this equation: it comes back throughout the course.
Path map
Detailed content
🧩 What it means to orchestrate multiple LLMs
The foundation: why a single model has a ceiling, what orchestration is, and the two questions that define everything.
An LLM is the AI model that runs behind ChatGPT, Claude, and similar tools—it receives text and returns text. Each model has its own strengths and weaknesses; none is best at everything.
Understanding that a single model has a “ceiling” is what motivates orchestration: if no model is best at everything, combining several may (in theory) go beyond any one model’s ceiling.
Single model · strengths/weaknesses by model · the ceiling of a solo model · the promise of combining.
Orchestrating means answering two questions: (1) who does each part of the work and (2) how the answers are combined Into a single result.
Every solution — Fugu, Fusion, Claude Code sub-agents — is just a different way of answering these two questions. It’s the map that organizes the entire course.
Routing (who) · fusion (how) · decomposition · aggregation.
The "maestro" (or conductor) is the one who reads the request, splits it into tasks, and assigns them to specialists — each one a model that’s good at that kind of task. At the end, someone combines the answers.
It’s the anatomy that repeats in any orchestrator. Recognizing the three roles (conductor, specialists, fusion operator) makes any "magical" system understandable.
Maestro/conductor · specialist · fuser/merge · pipeline.
The idea of bringing specialists together and calling on the best one for each case. When that’s packaged behind a single API, you hit one address and the orchestration happens behind the scenes.
That's exactly what Fugu sells. Understanding the packaging ("1 API, hidden team") explains why it looks simple on the outside and is expensive on the inside.
Mixture of experts · 1 endpoint · abstraction · hidden team.
When Claude Code opens sub-agents (haiku/sonnet/opus workers) and splits a plan among them, it’s orchestrating—just within one provider and under your direction.
Shows that orchestration is nothing new or exotic: you probably already use it. What changes in products is who decides and how automatic it is.
Sub-agent · worker · manual delegation · intra-provider orchestration.
Each specialist the maestro calls reads and writes tokens—and you pay for all of them, while waiting for each one to respond. More models = more time and a higher bill.
It's the counterweight to the promise. Without seeing the hidden cost, orchestration seems free; with it, you understand why Fugu came out 5× more expensive in the tests.
Orchestration tokens · added latency · overhead · cost × benefit.
🎚️ The spectrum: from hands-on to the maestro model
Who decides who does what? From “you put everything in one model’s hands” to “a trained model orchestrates on its own.”
Picture a slider: at one end, you write everything and decide who handles each part; at the other, a model decides all of that on its own. Products occupy different points on this slider.
Gives you a single map showing where to fit Codex+Claude by hand, sub-agents, Fugu, and Fusion—without memorizing each one in isolation.
Control spectrum · manual ↔ automatic · who routes.
You run Codex for one thing, Claude for another, and copy one model's output into the other's input. Full control, full effort—the orchestration is your manual work.
It's the baseline every product is compared against. Knowing the cost of doing it by hand shows what an automatic orchestrator actually saves you (and what it charges for that).
Manual routing · chaining outputs · control × effort.
Tools like Claude Code automate part of the delegation: you give it the objective, it opens workers and divides the work. The orchestration is automatic, but within a single provider.
It's where most people already are. It shows that "automatic" doesn't require cross-provider — and sets up the contrast with Fugu, which crosses providers.
Workflow · sub-agent · intra-provider · partial automation.
At the fully automated end, the logic of "who does what" lives in the weights of a model trained for this (Fugu). You send the request; it decides and routes across providers on its own.
It’s the new concept the course unpacks. Seeing that routing has become a "model weight," not your code, explains both Fugu’s appeal and its cost.
Routing in the weights · cross-provider · zero orchestration code.
Routing is the decision “who do I send this to?” It can be your rule, a workflow, or a small router model trained to choose the right specialist.
It's the heart of the "new way." Whoever masters routing masters the first of the two questions — and that's where the orchestrator's intelligence (and expense) lies.
Router/conductor · task-level decision · solve-it-yourself vs. delegate.
Moving to the right on the spectrum trades effort for convenience—but it usually costs more time and money because the automation calls more models than you would.
It gives you the criteria to choose the right point on the scale for each task, instead of adopting the most automated option just because it’s trendy. It’s the mindset the rest of the course explores in greater depth.
Control × convenience · cost × benefit · choose the right point for each task.