From Human‑AI Collaboration to AI‑Led Code Quality: Building a Digital SRE Agent for Automated Blocker Fixes

AI Involvement Evolution

Recent years have seen AI move from simple code completion and Q&A (e.g., ChatGPT, Tongyi Qianwen) to full‑task execution and even end‑to‑end development (e.g., Devin, AutoDev). The authors categorize this progression into four stages: initial assistance, IDE‑level assistance, AI‑native agents, and AI‑dominated development.

Challenges of Traditional Code‑Quality Governance

Late problem discovery : owners must manually log into quality platforms; no proactive alerts.

Lengthy, low‑standardized repair flow : locating the cause → branch checkout → local fix → code‑review (CR) takes 2–4 hours per application.

Reliance on personal expertise : repair quality varies with developer skill.

Governance loop difficulty : SRE teams spend most time on administrative tasks rather than strategic quality improvements.

Solution Overview: AI‑Led + Human‑Fallback Model

To address these pain points, the team built a digital SRE employee – an AI agent that autonomously performs Blocker detection, repair, and tracking, while humans only intervene at critical checkpoints (review and release).

“Construct a centralized ‘digital employee’ where AI handles Blocker discovery, repair, and tracking; humans audit and publish at key nodes.”

Core Capability Modules

Capability 1: Global inspection & task routing – a browser‑automation scraper logs into the quality platform, iterates over all applications, and extracts Blocker issues without modifying existing pipelines.

Capability 2: End‑to‑end automated repair – the agent parses issue details, generates patch code (including import handling), validates locally, creates a fix branch (e.g., fix/code-quality-blocker), pushes the change, and opens a merge request for the owner.

Capability 3: Transparent closure & human‑AI collaboration – an intermediate markdown log records each step (original description, file path, fix rationale, diff). This log is attached to the CR for audit and can be used for iterative natural‑language corrections.

Detailed Workflow

Repository discovery – resolve the repo URL from the application name.

Environment isolation – clone the repo and create a dedicated fix branch ( fix/code-quality-blocker).

Context analysis – read the scraped Blocker list, locate the offending lines in the codebase.

Intelligent repair – generate a fix, add missing imports, ensure no business‑logic regression.

Local verification – compile and run static checks.

Automatic PR creation – push the branch, open a code‑review request, and notify the owner via instant‑messaging.

If the owner finds the fix unsatisfactory, they can provide natural‑language feedback; the agent re‑runs the repair step with the new constraints.

Technical Trade‑offs & Choices

UI automation was chosen over direct static scanning (Error‑Prone) or CI‑pipeline integration because it required no changes to existing build environments and offered real‑time data without T+1 latency.

Two alternative paths (local static scans, CI data feed) were rejected due to high configuration cost and delayed data, respectively.

Pilot Results & Scaling

In an initial pilot of ~20 applications, the AI agent achieved a 60%+ automatic repair success rate and covered over 70% of Blocker issues. Real‑world feedback highlighted stability improvements (retry logic for dynamic pages) and robustness of the logging mechanism.

After two iterations, the system was rolled out to over a hundred projects, using a layered governance strategy:

Existing Blocker‑rich apps – fully automated repair with human review only at the PR stage.

New apps without platform integration – the agent assists in platform onboarding, performs an initial full scan, and brings the app into the unified quality view.

Efficiency analysis showed an ~80% reduction in direct labor cost compared with manual processes, and developers were freed from repetitive “syntax‑level” fixes to focus on higher‑value design work.

Future Outlook

The authors envision AI agents becoming "digital SRE" or "digital test experts," handling deterministic, repetitive tasks while humans define goals, set boundaries, and manage risks, establishing a new human‑AI collaborative development paradigm.