AI
What a Standardized Pipeline Actually Delivers
Nordstrom Standard Pipeline (NSP) runs automated builds, tests, and deploys for over 80% of Nordstrom’s services. Teams onboard by structuring a repo around a handful of config files. NSP hooks in and handles the rest: ingress, certificate management, base image patching, shared by domain experts. Developers inherit best practices without designing anything from scratch.
Loredana pulled six months of run data from one app before and after onboarding. Before: 136 CI/CD debug runs. After: roughly 10. That is a 93% reduction.
“the more that you lean on something like Nordstrom’s standard pipeline, the less you have to think about deploying” — Loredana Ordona
Rearchitecting NSP for GitHub Actions
GitLab generates a pipeline at runtime: the include keyword pulls shared templates into one large monolithic YAML file assembled behind the scenes. GitHub Actions favors smaller, composable reusable workflows. Nordstrom could not recreate the old model, so the team broke NSP into versioned, configurable components stored in a central repo. One engineer called them Lego blocks.
The shift is visible in the config files. GitLab: project-specific variables mixed with centralized template imports. GitHub: every called workflow is explicit and auditable. Teams still inherit the same guardrails automatically.
Why One Agent Failed and What Replaced It
Nordicist started with one agent handling everything: extracting GitLab variables, generating templates, populating values, opening the pull request. At that point, only 54% of code was AI-generated and only 28% of pull requests merged without manual commits. Prompting changes and new tools made no difference.
“We quickly learned one agent doing everything doesn’t scale well.” — Annabelle Watt
The fix was to split the work across two agents and add deterministic code where precision mattered. In the variable parsing and two-agent pipeline walkthrough (17:03): deterministic logic filters and scopes GitLab variables first. Agent one builds structure only, inserting placeholders where values are unknown. Agent two populates environment-specific values per workflow file.
Results, Hours Saved, and Lessons for Teams
After the multi-agent redesign, Nordicist generates 70% of migrated code and 40% of pull requests merge without any manual commits. Migration prep time dropped from four hours per repo to one, a 75% reduction. Across 5,000 repos on NSP, Nordstrom expects to recover nearly 15,000 engineering hours.
Annabelle’s three rules: split agent responsibilities so each piece is testable; treat prompts as product code, because every change shifts template selection or value population; keep a human in the loop before merging.
“Every prompt is part of your system logic.” — Annabelle Watt
Notable Quotes
the more that you lean on something like Nordstrom’s standard pipeline, the less you have to think about deploying Loredana Ordona · ▶ 7:56
We quickly learned one agent doing everything doesn’t scale well. Annabelle Watt · ▶ 15:43
Every prompt is part of your system logic. Annabelle Watt · ▶ 22:46
Key Takeaways
- Nordstrom cut CI/CD debug runs 93% after onboarding one app to its standardized pipeline.
- A single agent hit a ceiling at 54% code generation; splitting into two specialized agents plus deterministic logic raised it to 70%.
- Nordicist cut migration prep time from four hours to one, saving an expected 15,000 hours across 5,000 repos.