Triage failed GitHub Actions runs and self-heal flaky tests
Suneet Malhotra. Task: Suneet Malhotra deployed OpenClaw agents to automate CI failure triage at Motorola Solutions. The workflow monitors GitHub Actions, reads logs, classifies failures (flaky, regression, environment), and posts analysis to Slack/GitHub. A self-healing locator system attempts to fix broken selectors using accessibility trees and visual context. Engineers review agent outputs before merging changes, acting as the final arbiter for safety-critical decisions.; tools: OpenClaw. Human role: The original post describes the human role this way: Zero human coordination.; limitations: Production costs initially approached $4,000/month due to context bloat, requiring prompt optimization to reduce to $400/month. Agents require structured observability to track reasoning and decision points, not just success/failure. Hallucinations in high-stakes decisions necessitate confidence thresholds: auto-merge >95%, human review 80-95%, investigation <80%. Source reliability is secondary, based on a single practitioner's blog post..
Tools
Published on Agent Work Atlas:
Overview
Suneet Malhotra
Workflow
Suneet Malhotra deployed OpenClaw agents to automate CI failure triage at Motorola Solutions. The workflow monitors GitHub Actions, reads logs, classifies failures (flaky, regression, environment), and posts analysis to Slack/GitHub. A self-healing locator system attempts to fix broken selectors using accessibility trees and visual context. Engineers review agent outputs before merging changes, acting as the final arbiter for safety-critical decisions.
Reported result
The triage agent reduced average time-to-diagnosis from 45 minutes to 6 minutes within two weeks. Flakiness dropped from 8% to 0.8% due to the self-healing system. However, a hallucination incident occurred where an agent incorrectly modified assertions instead of locators, passing tests that validated wrong behavior. This was caught in code review, highlighting that agents cannot be trusted as final arbiters without human oversight.
Human role
The original post describes the human role this way: Zero human coordination.
Limitations
Production costs initially approached $4,000/month due to context bloat, requiring prompt optimization to reduce to $400/month. Agents require structured observability to track reasoning and decision points, not just success/failure. Hallucinations in high-stakes decisions necessitate confidence thresholds: auto-merge >95%, human review 80-95%, investigation <80%. Source reliability is secondary, based on a single practitioner's blog post.
Try it
I am a QA Lead at [replaceable company name] using OpenClaw. I need to set up a CI failure triage agent for our [replaceable tech stack, e.g., Playwright] tests. First, configure the agent to monitor GitHub Actions for failed runs, read logs, and classify failures as flaky, regression, or environment issues. Second, implement a self-healing locator system that suggests fixes for broken selectors using accessibility trees. Third, set up a confidence threshold: auto-merge fixes with >95% confidence, require human review for 80-95%, and flag for investigation below 80%. Before deploying to production, run a dry test on [replaceable sample failed test log] and show me the classification and suggested fix. Do not auto-merge any changes without my explicit approval. Measure success by comparing time-to-diagnosis before and after. If the agent hallucinates a fix that changes assertions instead of locators, stop and report the error.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.