← Projects
RUEN

AI agents across the task lifecycle

Client project

Agentic development at a large e-commerce company — from a product idea to production and docs

Problem

The team at a large e-commerce company closed 20–25 tasks a month, tech debt kept piling up, and routine work — specs, reviews, test deployments, QA checklists, documentation — ate up developers' time.

What I built

Three layers of agentic development: an agents plugin covering the whole task lifecycle, an “orchestrator + implementer” framework and a sprint coordinator. I'm the author of the approach and its implementation, as the team's lead developer.

1. An agents plugin for the whole task lifecycle

An internal plugin marketplace for Claude Code and Codex (macOS and Windows). 13 skills drive a task through the tracker statuses and team roles:

  1. Product idea → spec on a shared template, checked against the code and Swagger
  2. Review and estimate
  3. Capacity planning
  4. Development with tests
  5. Test deployment, migrations, triage of failed autotests
  6. QA checklist run — API and the UI in a browser
  7. Fixes from QA feedback
  8. Merge, backward-compatibility check with iOS and Android, production release, error monitoring in Graylog
  9. Documentation in the Wiki

2. An “orchestrator + implementer” framework

A stack- and platform-agnostic prompt for bringing agentic development into any project. A strong model writes the spec, the plan and reviews every commit; a cheaper model writes the code for one plan item in a subagent's clean context.

The plan is “spec-tight, code-light”: it pins down exactly what to do but holds no ready-made code, otherwise the savings disappear.

3. Sprint coordinator

Takes every sprint task and launches each one in a new session, driven by the orchestrator + implementer pair. Context doesn't pile up between tasks, and the strong model is spent on judgment only.

Engineering decisions

  • Integrations. Tracker and Wiki via MCP with pinned server versions; GitLab API, Graylog and SSH to the staging servers through my own Python CLI.
  • Secrets out of context. Secrets never reach the model's context or the chat: the CLI injects them itself, strips them from responses and never sends them to other hosts. 115 unit tests.
  • Human-in-the-loop. The agent first shows tasks, comments, pushes and status transitions and executes them only after confirmation; merges and production releases happen only on the release manager's word.
  • Hardened on real work. End-to-end runs of real tasks through all the skills; findings turn into plugin versions with a changelog for the team.
Result
  • 200+
    tasks run through the agents
  • 150–200
    tasks a month, up from 20–25
  • 90%+
    of tasks pass QA on the first try
  • Implementation from the spec needs no human involvement.
  • For the first time, the team is paying down tech debt systematically.
  • The Barcelona Design employee portal was built in the same AI-driven mode.
Every sprint task runs in a clean session, and a cheaper model writes the code