Experts

An expert is a reusable lens for one seat in a debate: "assume it fails", "you are the customer", "check the facts". A council made of experts is chosen by skills, not models: you name the experts you want, and the engine decides which model answers for each, spreading them across vendors and recording what it did.

What an expert is

An expert is a short written lens: what to look for, what to avoid, and which kinds of task it suits. It is stored as a standard Agent Skill folder, so it is readable, versioned and portable. An expert carries no tools and no permissions, and the text of a lens is shown in full before it is used. Eight starter experts ship in the engine, along with three presets of four experts each for common reviews: a risk review, a spec review and a search for gaps.

What is built

• Naming experts for a debate, or choosing a preset, in the open-source engine, with an opt-in setting.
• The engine assigns a model to each expert, spreading them across vendors, and records the assignment, with a hash of each expert, in the debate's receipt.
• Free, read-only tools that describe the starter experts to an assistant. They are off on this server.

What is planned, with no dates

• Running experts through the hosted server, with a price shown before you spend.
• More than one expert on the same model, which makes a council that is not multi-model. It is always labelled as such.
• A preview of the pairings and the cost before anything is spent, and a rule-based suggestion of which experts fit a question.

What has not been measured

Whether experts make answers better, and by how much, has not been measured. The cost model, the measurement plan and the rule that decides when a result may be claimed are in the methodology; its results are marked pending until something has been measured.