Agentic Workflow Audit — 2026-03-23

[github-actions[bot]]

Overview

22 workflow runs completed in the last 24 hours with a 90.9% success rate (20 success / 2 failure). Token usage was 12.0M tokens at a cost of $1.77. One failure is a known design-by-intent behavior; the other is a genuine concern with excessive agent looping.

---

Workflow Health Trend

Success rates have held steady at ~88–91% over the past week, recovering from the 72.7% dip on 3/15. Today's 90.9% continues the healthy trend. The 3/20 audit showed only 4 non-skipped runs (80% SR) due to low run volume that day.

---

Token & Cost Trend

Today's $1.77 spend is slightly below yesterday's $3.31 peak. Token volume of 12.0M is comparable to the 3/21 high of 13.1M. The 3-day moving average has risen sharply — driven by heavy Claude/Copilot usage in Documentation Updater and jsweep workflows.

---

❌ Failed Runs

1. GitHub Remote MCP Authentication Test — §23422628558

• Engine: Copilot (gpt-5.1-codex-mini)
• Tokens consumed: 3,289,458 (85 turns)
• Root cause: Agent entered an infinite loop of ● Listing background agents calls, never making progress toward the task. No safe output was produced, agent_output.json was never written, and the workflow concluded as failure.
• Impact: High — 3.3M tokens wasted with no useful output and no safe-output tool called.
• Recommendation: Investigate why the Copilot CLI agent loops on Listing background agents. This may be a background agent feature interaction with the MCP auth test environment. Consider adding a turn-count circuit breaker or a timeout safeguard, and verify the MCP server is reachable during agent startup.

2. Duplicate Code Detector — §23420898530

• Engine: Codex
• Root cause: By-design — duplicates were found, issue #22370 was created, and the workflow intentionally concludes as failure to alert maintainers.
• Impact: No real issue. The issue was filed with detailed refactoring recommendations for actions/setup/js/ expired-entity cleanup handlers.

---

✅ Notable Successful Runs

Workflow	Engine	Tokens	Turns
Daily Documentation Updater	Claude	3,642,117	52
jsweep - JavaScript Unbloater	Copilot	2,519,297	39
Copilot PR Prompt Pattern Analysis	Copilot	684,343	15
Issue Monster	Copilot	556,607	14

All 22 runs today

Run ID	Workflow	Engine	Conclusion	Tokens	Turns
§23419883184	jsweep - JavaScript Unbloater	copilot	✅ success	2,519,297	39
§23420748984	Smoke Claude	claude	✅ success	—	—
§23420748994	Smoke Copilot	copilot	✅ success	—	—
§23420748999	Agent Container Smoke Test	copilot	✅ success	143,742	4
§23420749027	Changeset Generator	codex	✅ success	—	—
§23420749036	Smoke Codex	codex	✅ success	—	—
§23420826663	Issue Monster	copilot	✅ success	556,607	14
§23420898530	Duplicate Code Detector	codex	❌ failure*	—	—
§23421309717	Daily Documentation Updater	claude	✅ success	3,642,117	52
§23421402437	Changeset Generator	codex	✅ success	—	—
§23421402463	Smoke Claude	claude	✅ success	—	—
§23421402467	Agent Container Smoke Test	copilot	✅ success	107,377	3
§23421402480	Smoke Copilot	copilot	✅ success	—	—
§23421402490	Smoke Codex	codex	✅ success	—	—
§23421408452	Safe Output Health Monitor	copilot	✅ success	—	—
§23422025128	Issue Monster	copilot	✅ success	423,015	11
§23422430166	Copilot PR Prompt Pattern Analysis	copilot	✅ success	684,343	15
§23422628558	GitHub Remote MCP Authentication Test	copilot	❌ failure	3,289,458	85
§23422824609	AI Moderator	codex	✅ success	—	—
§23422828677	Contribution Check	copilot	✅ success	321,476	8
§23422919303	Issue Monster	copilot	✅ success	319,821	8
§23423031288	Weekly Blog Post Writer	copilot	✅ success	—	—

* Intentional failure — duplicates found, issue filed

---

Recommendations

1. Investigate GitHub Remote MCP Auth Test loop — The Copilot agent spinning 85 turns on Listing background agents is a recurring risk for expensive no-op runs. Adding a max-turns guard or a health check before entering the main task loop would prevent token waste.
2. Monitor Daily Documentation Updater token usage — At 3.6M tokens / 52 turns, this is the heaviest consumer. Ensure the prompt scope is appropriate and consider if output can be constrained.
3. Duplicate Code Detector failure conclusion — Consider whether the GitHub Actions workflow conclusion should be success (issue filed = job done) vs failure (to draw attention). The current behavior creates noise in the success-rate metric.

References:

• §23422628558 — GitHub Remote MCP Auth Test failure
• §23421309717 — Daily Documentation Updater (highest tokens)
• §23420898530 — Duplicate Code Detector (by-design failure)

AI generated by Agentic Workflow Audit Agent · history

• expires on Mar 24, 2026, 6:03 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-03-24T06:03:44.281Z.

Closed by Workflow