Approach
Workflows before agents.
Most of the value is in choosing the right workflow and putting the right checks around it. The model is one layer of six.
Start simple
Anthropic's guidance for building agents is to find the simplest solution possible and only add complexity when it is needed. We follow it: a fixed workflow with one agentic step is often enough, and it is easier to test.
Source: Anthropic, Building effective agents (opens in a new tab)
The six layers we build
The AI model is the fuel. The harness is the engine that turns it into work you can trust. It gives the AI only the information about your business that it needs. It connects the AI to the systems it has to use, such as your inventory, your CRM or your machines. It follows your procedures step by step and tests its work on real cases before anything goes live. Before it spends money, signs anything or changes a customer's record, it stops and asks one of your people.
The brain
The AI model. We choose the one that fits your work and your budget, and we can swap it when a better one arrives.
Knowledge
What it may read: the documents, records and history it needs, and nothing else.
Connections
The systems it can use, such as your inventory, your CRM or your machines. It starts by reading; each new action is added on its own.
Playbooks
Your procedures, written down step by step, so it works the way your team works.
Quality checks
Tests built from your real cases. Every change runs them before it goes live.
Your approval
Before it spends money, signs anything or changes a customer's record, it stops and asks one of your people.
Risks we design for
Quality drift
An agent that passed last month can fail this month after a model, prompt or data change. Evals rerun on every change, and a person approves what matters.
Cost
Model usage is billed to your account, so you see every cent. Each agent has a hard monthly cap.
Data protection
You stay the data controller. The agent reads only what the permission map allows, in your cloud.
Vendor terms
Commercial API keys under your company's agreement with the provider, never personal subscriptions.
What we will not build
- Agents that decide on hiring, credit, housing or other matters about people.
- Customer-facing agents without a person in the loop.
- Anything that needs your data on our machines.
Open standards we build on
Model Context Protocol
The open protocol our connectors speak.
Agent Skills
An open format for versioned agent procedures.
AGENTS.md
One instructions file per repository that every coding agent reads.