When a team discovers agents, the first instinct is to want a single assistant: one dialog box you would ask everything, from pricing a contract to chasing a client. That is precisely what we do not build.
At Lilycare, eighteen agents run in production. Each one is narrow: one per insurer, one per portal, one per document type. None of them can do everything, and that is what makes them verifiable.
A narrow agent can be checked
An agent that does one thing exposes a clear contract: what it reads, what it produces, what it is not allowed to touch. When it fails, you know where to look. When it succeeds, you know what to generalize.
A generalist assistant fails silently: nobody can review a chain of reasoning that crosses ten trades. Trust is not declared, it is partitioned.
The graph makes the team
These agents are not isolated. They read and write the same data graph (contacts, companies, contracts, documents) under the same rules and permissions as humans.
Coherence comes from the graph, not from a central brain. Each agent leaves clean work behind, the next one picks it up, and the human team keeps its hands on the wheel at every step.