I designed and introduced a suite of AI agents implemented as Claude Code skills, configured to run automatically each day and integrated into the CompanyHouse engineering workflow. The objective was to incorporate AI agents into the complete software delivery process while retaining human responsibility for all material decisions (Human In The Loop).
A comprehensive Model Context Protocol (MCP) integration provides controlled access to the project repositories, read-only production databases, Elasticsearch logs, New Relic monitoring, Bitbucket and the Jira Scrum/Kanban board. Consequently, the agents can assess development tasks in their full technical and operational context.
The scheduled agent capabilities include:
- Production verification — daily first-pass testing of Jira tasks deployed to production, using database queries, error logs, HTTP checks and browser-based verification.
- Log-error triage — daily classification and prioritisation of production errors, including Jira duplicate detection and the preparation of corrective tasks.
- Technical pull-request review — daily assessment of new Bitbucket pull requests against Jira requirements, surrounding project code, security, correctness and performance.
- Refactoring analysis — daily review of rotating codebase areas to identify improvements supported by measurable business or operational value.
- Infrastructure monitoring — hourly verification of production-server health through New Relic, with Jira alerts for sustained threshold breaches.
- AI activity documentation — event-driven recording of AI involvement, human corrections and verification results in the relevant Jira task.
- Product scouting — daily research into competitors, market requirements and available CompanyHouse data to propose specialised product opportunities.
Claude Code was also introduced into the team’s regular development workflow, supported by explicit rules concerning permissions, auditability, verification and human ownership. Combined with CI process automation, this has shortened the delivery cycle for appropriately scoped tasks. Development, technical review, automated testing, manual verification and deployment can, in suitable cases, be completed within a single working day.