Compliance signed off on an agent for the first time, because the log was already the one they read

Harbourline runs card settlement for mid-sized merchants. Forty engineers, a regulator who reads their change logs, and a security team that had refused every coding agent so far on the same ground: nobody could say what it would do.

Where they started

Two earlier pilots had ended at the security review. Both agents could run shell commands; neither could show, in advance, the set of commands they would refuse, and neither produced a record the compliance team recognised as evidence.

What they changed

rovecode was different on two counts that had nothing to do with model quality. The deny-default rules are a file, reviewed and versioned like any other; security wrote them, and engineers could not loosen them locally once the organisation policy was in place. And every session is an append-only log with a hash chain — the same shape as the transaction logs the auditors already sample.

Enterprise rollout was mostly identity work: SSO against their provider, SCIM from the directory, IP allow-lists for the web app. Two days from a signed order to the first engineer signing in with their company account.

What they measured

Zero policy exceptions requested in the first quarter — the rules were right the first time, or close enough that the prompts covered the rest. Every session landed in the audit chain, because there is no way to run one that does not. The compliance review that had killed two pilots took an afternoon.

What was hard

The allow-lists. Engineers work from home on rotating addresses, and the first week the list was wrong twice a day. The fix was a VPN egress that had been on a roadmap for a year; the agent just made it urgent.