Backends for AI agents

• Java
• Spring Boot
• MCP
• REST

Role: Lead engineer

The problem

An AI agent is only as good as the tools and APIs behind it. After the first launch, a review of real questions found about twenty the agent answered poorly. The causes were not the model. They were missing tools and API limits the agent kept running into.

What I did

• Built the task, execution and remediation-planning APIs the agents call, including ready-made planning templates for common patching cycles.
• Turned the failure review into work: with the team I defined three new MCP tools, and proposed wrapper APIs for a query-size limit the agent kept hitting.
• Added validation-scan APIs for the launch.

What I took from it

• Treat agent failures like production incidents: collect them, group them, and fix the tooling, not just the prompt.
• Tool design matters. A tool that returns what the agent needs in one call beats one that needs five.

Details are kept general on purpose.