Teaching an Agent to Triage: Automating First-Level Build Failure Analysis

DRAFT — outline only; prose pending.

Alt title: "10,000 Pipelines a Month, Zero Humans on First Triage". Target: 2,200–2,500 words. Tone: engineering storytelling for senior readers, not a tutorial.

The toil problem

• ~300 words.
• 10K+ monthly pipeline runs; what a failed-build morning looked like; who got interrupted.
• One sanitized example of a cryptic failure that took a human 40 minutes.
• Scale numbers early.

Why "just point an LLM at the logs" fails

• ~400 words. Credibility section.
• Context windows vs multi-MB logs; the agent must navigate Jenkins, not read a dump.
• Hallucinated root causes without build-system context.
• Triage is a procedure, not a prompt.

The architecture: tools + skill

• ~600 words + diagram.
• The 11 Jenkins REST API tools and the design principle: give the agent the same affordances a human triager has (get build status, fetch logs incrementally, list recent builds, diff against last green…).
• The SKILL.md: the failure taxonomy (compilation, test, dependency, infra, deployment, timeout, permission, quality gate) and the encoded triage procedure — the "imagination layer": writing down the expert's implicit decision tree.

What structured output looks like

• ~300 words.
• Sanitized root-cause report: category, evidence excerpts, confidence, recommended action, handoff owner.
• Before/after: cryptic red X → ticket-ready summary.

Results and honest limitations

• ~400 words.
• % correctly categorized, human-minutes saved, MTTR movement.
• Where it still fails (flaky tests, novel failure modes); what you'd redesign.
• The honesty is what earns senior trust.

The generalizable lesson

• ~250 words.
• Pattern = codify expert procedure as a skill + give the agent real tool affordances.
• Applies to incident triage, capacity review, dependency upgrades.
• One line teasing the series ("build triage was the first skill; code search came next").

Pre-publish checklist: sanitize repo names/hostnames; check employer sign-off (an Auction.com eng-blog byline + cross-post is the best case — free amplification); title = honest + specific + numbers.