You need a research agent by Friday. Instead of starting from an empty file you clone a template, drop in one key, and a quarter of an hour later something on your screen is searching the web, reading pages, and writing a summary with the sources listed underneath. It runs. The temptation is to feel most of the way there, and in one sense you are: the shape is right and you did not have to invent it. But nothing you just downloaded knows which sites your company is allowed to read, or what should happen when the summary turns out to be wrong.

A proven hull settles a hundred decisions

Two kinds of starting point are on offer, and they are not the same thing. One is a whole runnable application for a single use case, with the model, the loop, the tools, the memory, the interface, the configuration and the dependencies already wired to one another. The other is a shelf of worked examples, each showing a single technique on its own: an orchestrator handing pieces of work to specialised workers, a sub-agent, several tool calls made at once, a long history compacted so it still fits.

Think of a boat. The first kind is a hull that has already crossed an ocean, drawn plank by plank, with the rig and the ballast sized against each other — a taller rig needs more ballast, more ballast wants a different hull, and the drawing settled all of that before you arrived. You could work it out yourself, over years, mostly by being wrong. The second kind is the boatyard’s book of details: how a joint is scarfed so it never opens, how a fitting is bedded so it stays dry. Each detail is shown alone, correct, and silent about where it belongs. Both are worth more than a description, because both come with real dimensions instead of adjectives. Reading that an orchestrator delegates to workers leaves you a dozen decisions of ambiguity. Running one shows you what the harness does.

What you inherit is not only the speed

A hull that has crossed an ocean crossed someone else’s. The drawing carries their weather, their crew, their harbour. A template carries the same freight: credentials that belong to a demonstration, tools left wide because wide is convenient when nothing is at stake, and just enough logging to make the run look good on a recording. All of it becomes yours the moment the thing goes in front of a customer, and none of it announces itself.

The team that starts from the research template and finishes well does unglamorous work on top. It puts in its own retrieval provider, restricts which domains may be read at all, keeps the trace of every source so a claim can be walked back, and sets a quality gate in front of anything that reaches a client. That is not the template failing. That is the part a template was never in a position to do. End-to-end describes coverage, not verification, and what arrives is code to be hardened rather than an architecture already proven anywhere but on the author’s machine. A single worked example carries the same debt: copy it without redesigning permissions, budgets and failure modes for your own domain and you have a shape that runs and a threat model that does not exist.

A pattern travels better than its reasons

Something genuinely useful is forming here, which is a shared vocabulary. Orchestrator-worker, sub-agent, parallel calls, compaction: each of these now points at something specific, because runnable code sits behind it and closes the gap between how a design is described and what a harness does once it is running. That is a real gain, and it is why a system of several agents is easier to talk about today than it was a year ago.

The old failure comes along with it. What spreads is the shape, because the shape is short and the reasoning is long, and the reasoning is where the choice lives. Reach for an orchestrator and workers because your task really does split into parts that do not need each other, not because it is the arrangement you have heard about most. The one that fits is often the plain one, and you find out which by measuring your own problem: how much it carries, what it may touch, what it costs when it is wrong. That measuring is the work nobody hands you in a box, in the same way that the loop, not the model, is where the behaviour you care about is decided.

A proven design gives you the hull. The sea it has to survive is yours.