The harness

The model is the engine. The harness is everything that turns it into an accountant: the loop, the context, the tools it's given and the rules it works under. Finn is the harness.

The agent loop

Finn plans, calls a tool, reads the result, and decides what to do next, until the job is done. Ask "have I paid everyone this month?" and she'll check bills, match payments, look at what's overdue and come back with an answer and the next step, not a guess.

Thinking before answering

Finn reasons before she replies and between tool calls, so a multi-step job (reconcile, then categorise, then prepare the return) is worked through, not rushed.

Context engineering

Every turn, Finn loads what she needs about your business: your company memory, upcoming deadlines, a money summary, recent activity and the page you're looking at. The stable part of her instructions is cached, so replies are fast and the cost per turn stays low.

A lean, tuned system prompt

Finn's core instructions are kept short and are retuned as models change. Rules that belong to one job live with that job's tool or in her knowledge base, not piled into one giant prompt. That keeps her sharp as the product grows.

Tool search

Finn has 255 tools. Loading all of them on every turn would crowd her context, so the everyday ones are always there and the rest are found on demand, the moment a conversation needs them.

Capability gating

Finn only gets the tools that fit your company. A UK sole trader never sees a US payroll form, and a Canadian company never sees a Companies House filing. The rules for which country and entity gets which feature live in one registry, checked before every tool call.

One Finn, every channel

The same agent, with the same memory and tools, answers you on the web, by voice, by video, on the phone, on WhatsApp, Telegram, Slack and email, and from inside ChatGPT and Claude. Start on your phone, carry on at your desk.