MCP servers, Claude Code plugins and agent tooling

The problem

Agents are only as useful as the tools they can call and the guardrails around them. The team needed agents that could do real engineering work (reviewing PRs, generating code, editing many files at once) and a platform that showed what those agents and training jobs were actually doing.

Approach

• Built and maintained Claude Code plugins, 86 skills across 6 plugins, for the engineering team.
• Built MCP tool harnesses that power AI-assisted workflows: autonomous PR review, code generation and multi-file editing across the org.
• Built a tournament system with daily reporting cronjobs, error-log surfacing and diagnostic scoring to evaluate multi-agent RL policies.

Architecture

• Observability pipeline on Datadog and OpenTelemetry (OTLP receivers, structured collectors) that replaced legacy monitoring.
• A Kubernetes watcher service for AI job lifecycle orchestration, with automated failure classification (timeout, OOM, policy error), stuck-job reconciliation and keep-main-green CI.
• Terraform for AWS (Lambda, Secrets Manager, IAM) behind the GitHub webhook integrations and deploy pipelines.
• A 21-table PostgreSQL schema with Alembic migrations and pre-deploy Helm jobs for zero-downtime releases.

Results

• Agents took on PR review, code generation and multi-file edits for the engineering team.
• Job failures were classified automatically and stuck jobs reconciled without manual work.
• 1,180+ commits shipped to the open-source Metta AI platform.

Stack

• Claude Code
• Model Context Protocol
• Python
• Kubernetes
• Datadog
• OpenTelemetry
• Terraform
• PostgreSQL