The problem
Some questions call for more than a quick search. “Compare train, plane and car for a London to Edinburgh trip next month, prices and travel times included.” “Give me an overview of the e-bike market.” The main assistant knows how to search, but it's juggling your memory, your tools and the conversation at the same time. A specialist focused on a single task, with a precise role, often does better.
But delegating has a cost: each expert makes its own calls to the model. It had to be worth it, and it had to stay under control.
What I wanted
I wanted an ephemeral expert: it exists for the length of one task, works within a strict frame, then disappears, leaving you a written analysis. You don't talk to it directly. You talk to LIA, who organizes the delegation and presents the result to you.
How it works
LIA delegates when your request really benefits from a specialist's eye: an in-depth analysis, a comparison that crosses several sources, or several independent points of view, like a financial expert and a legal expert on the same subject. Simple questions are still handled directly. You can also ask for it yourself: “answer me like an analyst, use a specialized sub-agent”.
- A role and a brief. LIA gives the expert a role, that of a specialist in the field, and a task written for your question, with the useful information from your request already included.
- A few tools, read-only. By default, web search and page reading. It sees neither your e-mails, nor your calendar, nor your contacts: whatever LIA doesn't pass on in the brief, it doesn't know. The administrator chooses the list, a tool that would change anything isn't allowed in it, and an expert can't launch another one.
- A method. It separates what it observed from what it infers, dates its figures, cites its sources, and says clearly when a piece of information is missing. The text of the pages it reads is treated as data, not as an instruction to follow.
- Limited time. It has a limited number of search rounds. When they're used up, a final pass writes up what it found.
- Several at once. Independent experts work in parallel, and LIA presents their analyses to you.
The analysis reaches you as the expert wrote it, with its sections and its sources: LIA doesn't cut it down to two lines.
What it costs, and who decides
An expert costs several times as much as a direct answer, and several experts in parallel cost even more. What they consume is added to the cost of your request, counted and displayed like the rest. In ReAct mode, LIA asks for your approval before launching an expert. The feature itself is turned on or off for the whole installation, by the administrator: it isn't a setting of your account.
The limits
An expert only reads and analyzes: it sends nothing, changes nothing, books nothing. Its web search goes through the Brave or Perplexity keys you've connected; without them, it can only read pages whose address it already knows. Its analysis is only as good as the sources it finds, and nothing guarantees it has seen everything. Finally, deciding to delegate remains the model's judgment: it may sometimes call in an expert where a direct answer would have been enough.
Read more: Pipeline or ReAct: two ways for LIA to handle your request; A model for each task, from the provider of your choice.