How Finn works
Anatomy of the agent
The parts that make up Finn and how they fit together.
A good agent is more than a good model. It's six layers working together, and every one of them has been built for accounting.
| Layer | What it is | In Finn |
|---|---|---|
| Model | The reasoning engine | Claude Sonnet 5.5 by default, with other models for specific jobs, swappable |
| Harness | The loop that drives the model | Finn: plans, calls tools, checks results, acts again |
| Tools | Actions the agent can take | 255 tools on real books |
| Memory | What it remembers | Conversations, company memory, personal preferences, team context |
| Knowledge | What it knows and trusts | Official sources, country playbooks, a verified rules library |
| Guardrails | What keeps it honest | Approval before action, checked replies, exact maths, a full log |
Underneath all six sits the system of record: a double-entry general ledger. It's the one thing a general-purpose AI can't bring to your accounts, and the reason Finn's answers are numbers you can file.
The model is the engine. Finn is the driver.
Models change every few months. The harness, the tools, the memory and the ledger are what make Finn useful, and they stay put when the model underneath gets better. When a stronger model comes out, Finn gets stronger the same week.