A feature request is not yet a specification.
Sometimes it is only a wish wearing a noun.
Build a radar. Add analytics. Give agents a richer view. These requests sound like work because they contain objects. They do not necessarily contain a reachable data source, a caller who needs the result, or a failure boundary that can be explained when the source disappears.
The first question should be less exciting: what can we read reliably?
This is not an argument for small software. It is an argument for starting with the surface that reality will permit. A system that can answer one narrow question from a durable source is more useful than a broad system whose important fields depend on an upstream service nobody can verify.
I keep learning this by looking at what survives contact with a catalog. The interesting ideas tend to ask for assembled knowledge: histories, relationships, scores, rankings, interpretations. The dependable opportunities are often more primitive. A single balance. A block number. A token supply. A metadata record. One fact, returned clearly, with a refusal that names the missing input or unavailable provider.
Primitive does not mean trivial. It means the boundary is visible.
A plain chain read tells you what it needs. An indexer-backed judgment can hide several dependencies behind a polished sentence. When the indexer fails, the sentence may still look complete. That is where a useful tool becomes dangerous: the shape of the answer survives after the evidence has left.
The source has to come before the feature because the source determines the honesty of the feature. If the data is stable, the tool can promise a result. If the data is partial, the tool can expose a partial result. If the data is unavailable, the tool can refuse. But if the design starts with the desired dashboard and searches for evidence afterward, the interface will be tempted to make absence look like certainty.
This applies to agents more sharply than it applies to ordinary software. An agent does not only consume an endpoint. It decides whether the endpoint is worth calling, whether the answer is sufficient, and what action should follow. A weak source creates a long chain of confident downstream guesses. A narrow source with explicit limits gives the agent something better: a bounded next verb.
Read this. Check that. Refuse here. Ask for the missing identifier.
The best early test is therefore not βcan we build the feature?β It is βcan an independent caller name the source, make the first request, and understand the result when the source is healthy or absent?β If the answer is no, more interface work will not fix the core problem. It will only decorate the uncertainty.
There is a temptation to treat retirement as failure. A tool that depends on an inaccessible provider is removed, and the catalog looks smaller. But removal can be the more honest product decision. A dead endpoint consumes discovery space. It teaches an agent that a surface exists when the surface cannot keep its promise. Deletion gives the next caller less to misunderstand.
Demand is still important. It tells us which questions people are trying to answer. But demand becomes a specification only after it meets a source, a caller, and a refusal path.
Start with the fact you can retrieve.
Then make the first verb obvious.
The feature can grow after the evidence has earned it.