Claude Code, a key piece in building a multi-agent platform that audits cloud spend

Leroy Merlin needed to review on an ongoing basis how much it was spending on the cloud and how it was using its resources, with a view to optimising them. Until then, that audit was carried out manually by the internal Operations team, going project by project and drawing up recommendations according to each person's judgement. The process worked, but it came at a high cost, consumed many hours of work each week and depended on the subjective judgement of whoever carried it out. The challenge was not only to save the team time, but to turn that manual analysis into an automated, recurring system that fitted into the team's daily operations.

Diverger built a platform that took on part of the Operations team's work, in which the system analysed the metrics and consumption patterns of each cloud project and generated, on that basis, corporate reports with prioritised recommendations, a concrete action plan and an estimate of return on investment (ROI). All of it aligned with Leroy Merlin's processes and technology policies.

In practical terms, Operations staff could obtain a complete report on any cloud project in a short time and, from it, take sound decisions to optimise their resources.

To achieve this, the system needed judgement comparable to that of an Operations manager, so a platform of specialised AI agents was designed, able to analyse, assess and write each report. It also had to integrate with Leroy Merlin's corporate tools and include its own evaluation system to verify at all times that the reports were reliable and of good quality.

Building all of this to a tight deadline and with a small technical team was no simple matter. For that reason, Diverger decided to bring Claude Code into the work, not at one particular moment, but across every phase of the project.

In building the multi-agent system, Claude Code helped the team develop more than ten AI agents, coordinated with LangGraph and running on Argo Workflows in Kubernetes. During testing, it helped write the automated tests and simulate external tools such as Datadog, GitHub and MLflow, so the team could validate the system without depending on those external services for every test run. The project reached more than 750 automated tests, a figure hard to achieve in the time available without that support.

In integration and security, Claude Code sped up work such as implementing more than 40 endpoints for communication with other systems, developing the gateway that handles access using Leroy Merlin's authentication, and designing a permissions model in which each agent reaches only what is strictly necessary for its task. The team also drew on Claude Code to write all the project's technical and functional documentation.

With Claude Code taking on most of the implementation, the testing and a first pass of code review, the technical team could concentrate on deciding what to build and why, rather than on mechanically writing every line of code.

The project therefore started from a clear objective and demanding conditions: delivering a technically complex platform, to a fixed deadline and with a small technical team.

With Claude Code's support, Diverger delivered what Leroy Merlin expected — a platform in production with more than ten AI agents, 40 endpoints and 750 automated tests, backed by complete technical and functional documentation, on time and to the quality agreed. But the impact went beyond meeting the original objective: by freeing the technical team from the more mechanical work, Claude Code made it possible to invest that time in designing a more robust and scalable platform than planned. What began as a tool to automate one specific Operations task ended up as a corporate solution ready to keep growing.