Configured once. Inherited everywhere.
Every chat, sheet, and workflow opens already knowing your model, your memory, your sources, and your rules — and works under them.
Hands, not chat
The agent works through the product’s own actions — it builds sheets, attaches documents, runs workflows. If the UI can’t do it, the agent can’t either.
Plan, then execute
A big ask produces a plan first. Approve it, and every step runs in an open trace you can watch and stop.
You hold the dial
Four modes — Ask, Plan, Accept, Auto. Reads are free, writes are gated, deletes always ask — in every mode, including Auto.
Deletes always ask.
Memory, scoped like a deal room
The agent remembers what you tell it — per user and per project, hard-isolated between deals, dated ‘as at’ so a value holds at its committee date.
Guarantor covenant tested annually — as at 12 Jun 2026
Context on tap
Answers ground on the deal’s documents and six live UK data sources — the grid queue, the renewables register, Companies House, the planning register, the six regional network records, and the European transparency platform — on your own credentials, with every call logged.
Any frontier model
Claude and OpenAI side by side, reasoning tiers per task, one-line swap. The model is not the moat.
Everything on the record
A read-only chronicle of what the agent did and which sources it called — the audit trail a regulated buyer expects.
Skills & Templates
The same spine will carry firm-specific skills and real Office templates — your house IC-memo format, your rulebooks — so the agent works your way, not a generic way.