LangChain Engineer Builds Self-Healing Agent Pipeline with Automated Regression Triage
Summary
- • LangChain engineer built a self-healing pipeline that auto-detects and fixes post-deploy regressions
- • Open SWE coding agent writes fix PRs with no human intervention until review time
- • A 7-day error baseline vs 60-minute post-deploy window separates real regressions from background noise
- • Two distinct paths handle Docker build failures and server-side runtime regressions separately
Details
Self-healing GitHub Action triggers on every deploy to main
The Action fires after each push to main, capturing build and server logs and routing them to one of two triage paths before handing off to the coding agent for remediation.
Docker build failures: CLI logs + git diff passed directly to Open SWE
If the Docker image fails to build, the pipeline pipes error logs and the latest commit diff to Open SWE, which researches and authors a fix PR with no human involvement — build failures are almost always caused by the most recent change.
7-day error baseline established with regex normalization
Error logs from the past 7 days are normalized into signatures — UUIDs, timestamps, and long numeric strings are replaced via regex and entries truncated to 200 chars — so logically identical errors bucket together regardless of dynamic values.
60-minute post-deploy window polls errors from the current revision
After deployment, the system applies the same normalization and compares the new error set against the baseline to isolate regressions genuinely caused by the change rather than pre-existing noise.
Stack: Deep Agents + LangSmith Deployments + Open SWE (open-source)
The GTM Agent runs on Deep Agents; LangSmith Deployments handles the deployment layer; Open SWE is an open-source async coding agent capable of codebase research, fix writing, and PR creation.
Author: post-deploy triage is harder than shipping code itself
The engineer frames the motivation as wanting to deploy and trust the system to catch problems — positioning autonomous triage as a team velocity and reliability problem, not just a technical optimization.
New Tech = new capability or system; Tech Info = implementation detail; Infrastructure = platform or stack component; Insight = author framing or conclusion
What This Means
This architecture offers a concrete, production-tested template for closing the deployment feedback loop using agentic systems — automating regression detection and fix authoring to reduce on-call burden for teams shipping AI agents. The error normalization and baseline-comparison technique is the key design detail that makes autonomous action trustworthy rather than noisy.