What goes into a corporate agent harness

We build internal AI systems for companies. Most of our work is the layer around the model, which we call the harness. Here is what is in ours.

Agents run with one employee's rights

An agent gets the access rights of the employee who gave it the task and sees nothing that employee can't. It works in a closed environment, and its outbound traffic passes through a filter. Passwords and keys to the customer's systems are stored encrypted.

Personal data is masked before it leaves

The checks run inside the customer's perimeter. Tax IDs, IBANs and card numbers are verified by checksum, and a separate detector finds names. Before a request goes to an external model, the values found are replaced with placeholders, and they are put back in the finished response. The mapping table is encrypted. If masking fails, the request is not sent.

The model is chosen per task

Models run at external providers, on rented compute or on the customer's own hardware. Large models handle reasoning and analytics. Small local ones handle high-volume processing and sensitive data. When a new model ships, the customer decides whether to switch.

Every request is logged

The log records logins and AI requests: who asked, which model answered and what it cost.

Connectors and modules

The harness connects to what the company already runs: SAP, NetSuite, QuickBooks, email, CRM. Each customer has their own modules on a shared core. We start with one module as a pilot, measure the result, then bring in other departments.

Two costs that are easy to miss

Security has to be rethought from the ground up. Access rights were designed for people, who read one document at a time. An agent with the same rights reads all of it in minutes, and every email it reads can instruct it.

The assistant answers from documents, and the documents are older than the practice. Things people fixed informally years ago come back as the official answer.

The product page is https://os.tenzorro.com/en