Carbon-Aware Test Scheduling in Agentic CI/CD: Risk-Proportionate Agent Tier Routing and Grid-Intensity Deferral

This paper introduces a carbon-aware test scheduler for the K11tech Agentic AI QA System that reduces CI/CD pipeline carbon emissions through two complementary mechanisms: risk-proportionate tier routing, which maps PR risk score to agent tier (CHEAP, MEDIUM, or FULL) to eliminate unnecessary full-suite execution; and grid-intensity deferral, which schedules expensive agents to the lowest-carbon time window in the next six hours using the Green Software Foundation Carbon Aware SDK.

Evaluation on 180 real pull requests demonstrates that tier routing reduces per-PR CI carbon by 52.5% for low-risk PRs and 38.7% overall; grid-intensity deferral reduces carbon intensity by 52.4% for full-tier agents, saving 16.80 g CO₂ (31.4% of baseline) for HIGH/CRITICAL workloads. Both mechanisms preserve defect detection quality (F1 = 0.931). Deferred jobs wait a median of 2.1 hours.

The scheduler is validated through a five-phase test suite of 197 tests covering unit, integration, agentic, end-to-end, and carbon accounting layers. It integrates with any agentic QA pipeline via a zero-modification adapter pattern and is released as open-source software under the Apache 2.0 license. This work aligns with the EU Corporate Sustainability Reporting Directive (CSRD) and the Green Software Foundation Software Carbon Intensity (SCI) specification.