How we build

EDITH is the operating layer our products are built on. It is not a model and not a chatbot. A model generates language; EDITH decides what has to happen, delegates the work to the right models, agents and tools, runs it, stops where the authority rules require a human, and keeps the record.

Because the layer is above the models, the models are replaceable. That is a design requirement rather than a slogan — providers retire models on short notice, and a system that has to be rewritten every time one does is not a system you can rely on.

Actions are typed by class. The classes that matter — privileged, destructive, external, personal-data, and AI-initiated — do not proceed without explicit owner approval. Each approval is cryptographically fingerprinted to the exact command and parameters it authorises, and is single-use.

Alongside that, capability grants mean an executor cannot perform what it was never granted. The two mechanisms are independent on purpose: one limits what may be attempted, the other limits what may proceed.

Every mission leaves a durable, append-only account of what was planned, what was approved, what executed and what resulted. When something is interrupted mid-flight, mutating work fails closed for human review rather than being silently retried.

Approval gating and trace emission are increasingly available as managed services, open-source features and free tooling. We do not present them as our moat.

What we think remains genuinely hard is making the authority model explicit and correct per capability, keeping execution fail-closed under adversarial input, and producing evidence that an auditor will accept across a heterogeneous set of vendors. That is the layer we are building, and we would rather concede the rest than overstate it.