The challenge
Northwind, a London-headquartered B2B SaaS company in the workforce-analytics space, had 180 engineers across four product lines. Code review queues averaged 38 hours, refactor work was perpetually deprioritized, and senior engineers spent roughly a third of their week on glue work — writing tests, updating docs, migrating between framework versions. Leadership wanted measurable throughput gains without expanding headcount or compromising their SOC 2 Type II posture.
What we built
A staged rollout of agentic coding capability across the org:
- A Claude Code-based agent harness wired into Northwind's monorepo with scoped read/write permissions per team, gated by their existing GitHub Enterprise SSO and CODEOWNERS rules.
- Custom skills for the patterns that mattered most: writing failing tests first, generating ADRs from PR diffs, performing dependency-bump migrations across 240+ services, and authoring runbooks from incident postmortems.
- A reviewing-agent layer that runs on every PR — checks against the internal style guide, flags untested code paths, and posts inline suggestions before a human reviewer is ever paged.
- An MCP server exposing Northwind's internal API catalog, design tokens, and customer-tier feature flags so generated code conforms to house conventions on the first try.
Results
Across 11 months: 15,000 engineering hours recovered, PR cycle time down 61%, and shipping velocity (story points completed per sprint) up 2.3x. The platform team retired two long-stalled migration projects in a single quarter. Crucially, defect-escape rate fell 22% — agentic review caught issues that previously made it to staging.