How we build Fyord.

Fyord is an autonomous AI employee operating on real customer money, real customer relationships, and real warehouse inventory. The bar for reliability is high. Here are the principles we work from.

1. 01

   Best-in-class providers and state-of-the-art models.

   We don't pin to one vendor or one model. Fyord uses Anthropic's Claude family — Sonnet for reasoning-heavy work, Haiku for cheap-and-fast triage — and we revisit model selection as new releases land. Our orchestration layer abstracts the model so we can swap upward as the state of the art moves.

2. 02

   Investing in agent runtime and harness.

   Naïve LLM calls aren't enough. Fyord runs on a proper agent harness — structured tool use, retry semantics, deterministic state machines around the non-deterministic parts, and a clear separation between agent reasoning and side-effecting actions. Failures are observable; retries are bounded; nothing happens twice that shouldn't.

3. 03

   Mechanisms for agents to learn your business.

   A generic AI doesn't know that BG-4417 is the same as “regel 45×95” to your customer Ola. Fyord's knowledge base — customer-specific aliases, learned product mappings, training pairs from your past orders, your custom checklist rules — gets richer every week. The agent doesn't get smarter at wholesale in the abstract. It gets smarter at YOUR wholesale.

4. 04

   Guardrails and loops.

   Every action the agent takes is gated. Pre-execution: does this order pass the customer's checklist? Are line items recognized? Is the customer resolved? Post-execution: did Fortnox accept the payload? Did side effects match expectations? When in doubt, the agent doesn't push — it asks. Bounded retry loops with backoff handle transient failures; persistent ambiguity escalates to a human.

5. 05

   Reliability through layered architecture.

   A single agent making every decision is fragile. Fyord uses specialized sub-agents — one for classification (is this an order?), one for extraction (what are the line items?), one for clarification drafting (how do we ask the customer?). A supervisor coordinates. Pre-processing flows handle structured (Excel, structured PDFs) and unstructured (free-text emails, scanned attachments) data differently. LLM-as-judge verifies critical outputs. Human-in-the-loop is built in, not bolted on — for the first weeks of every customer's onboarding, every order is reviewed before it lands in Fortnox.