Owning a capability decides who is answerable. On its own it does not put the facts where anyone can reach them, record why a call was made, route a question to the team that owns it, or build the team that does the work. Those are the five the home page sets out, and the argument behind them is in the notes. This page is the machinery they run on. Some of it you will already have, and the useful question is not what a greenfield build would look like, it is what the gaps are costing you.
Read down the list. Stop at the first one you answer no to, and follow the link on that row to see what is listed there. It is a list rather than a review: nothing on it is recommended, and a product missing from it was not rejected.
You have it ifAn agent run gets a workspace that is not somebody's laptop, and two runs cannot see each other.
Empty, it costs youEvery unattended run inherits one machine's credentials and state. This is the row that stops being optional first, because a scheduled job on a laptop stops when its owner takes leave and nothing reports that it stopped.
You have it ifYou run a process that holds weights, or you have decided on purpose that you never will.
Empty, it costs youNothing, if you buy inference and mean it. Empty is the correct state of this row for anyone who does.
You have it ifChanging which model answers is a configuration change rather than a code change.
Empty, it costs youEvery switch becomes a code change in however many places call a model. One model and one consumer does not earn this row. A second of either does.
You have it ifAn agent reaches your systems through a declared surface rather than through pasted text.
Empty, it costs youNothing records what an agent was actually given, so when a run comes back wrong there is nothing to inspect and nothing to change except the next prompt.
You have it ifA sequence that dies halfway resumes rather than restarts.
Empty, it costs youLong work is only as reliable as the machine it started on. This row earns its place once a run is long enough that starting it again is expensive, and before that it is easy to buy too early.
You have it ifNobody has to read a log to know what an agent did.
Empty, it costs youYou cannot tell a working agent team from a stuck one without going and asking the person running it.
The order is the order these stop being optional in, which is an argument about what depends on what and not about what is good. Five groups in the options have no row here, on purpose. Skills and specs, and security and secrets, are not on the request path in the way these six are, so an order of adoption says nothing useful about them. Testing agent output is not a row either. You can buy what runs an evaluation, and the options list seven products that trace, score and test, but none decides what a good result is: those cases are yours to write. Nor is distribution to teams, where the options hold one vendor feature. Catalogs and registries are not a row here either, and that is not scarcity: they have rows in the options. What it does not have is a place in this order, and this page does not invent one. Naming that is the alternative to selling you a gap as a layer.
None of this has been measured against an organization, and the ordering is reasoned from what depends on what. If a row is wrong, or the order is wrong where you are, say so, and today I read it myself.
All the options, in one table ·Read the notes ·How the work gets done ·Crinaro.AI