Cost, Context & Efficiency Tradeoffs

Several chapters in this course have touched a recurring theme in passing — orchestration has a real cost, worth weighing against its benefit. This chapter makes that theme the explicit main topic: what orchestration actually costs, what it genuinely buys, and how to decide when the tradeoff is actually worth it.

What Orchestration Actually Costs

Running multiple agents instead of one means more total model calls, more overall tokens processed across the whole workflow (since each agent needs its own briefing and context), and — in a sequential chain (Chapter 3) — more total wall-clock time than a single agent doing the whole task itself. These are real, measurable costs, not abstract concerns to wave away.

Context Costs Specifically

Fundamentals' Chapter 1 named a genuine context-efficiency benefit of delegating to a subagent: large, disposable exploratory work stays out of your own main conversation. But this course's own Chapter 2 (programmatic definitions adding maintenance complexity) and Chapter 7 (shared state risking its own bloat) both showed real costs on the other side of the ledger. The honest picture is genuinely two-sided — not "agents are always more context-efficient," but a real tradeoff to weigh in each specific case.

When Orchestration Pays For Itself

Fundamentals' own capstone closed with a direct point worth repeating here: the real payoff of formalizing a full agent team or pipeline comes specifically from doing that work at real, repeated volume, where a fixed setup cost gets amortized across many uses. Building an elaborate pipeline for a task that will genuinely only happen once or twice rarely earns back its own setup cost.

When It Doesn't

For a genuinely simple, one-off task, doing the work directly — without launching any subagent at all — is often both faster and lower in total overhead than briefing and launching a separate agent for it. Delegation is a tool reserved for tasks that genuinely benefit from isolation, parallelism, or specialized focus — not a default reflex applied to every request regardless of whether it actually needs any of that.

Parallel Agents and Cost

Chapter 3 established that parallel execution reduces total wall-clock time — but that speed benefit doesn't come free. Running several agents at once roughly multiplies the total resource cost by however many agents are running simultaneously, even though the elapsed time is shorter. Parallel orchestration is a genuine time-vs-total-cost tradeoff, not an unambiguous win on every dimension at once.

Hands-On Exercises