[audit-workflows] Daily Agentic Workflow Audit — 2026-04-12

[github-actions[bot]]

Automated daily audit covering 85 workflow runs from April 12, 2026 (16:00–21:04 UTC). Overall health is good with an 88.1% success rate (37 successes, 5 failures, 39 skipped). Token usage and cost rose significantly day-over-day: 26.5M effective tokens and $15.73 today vs. 8.5M tokens / $8.44 yesterday, driven by more active PR/discussion triggers and the Q workflow hitting a 10M-token run.

Summary

Metric	Value
Total Runs	85
✅ Success	37
❌ Failure	5
⏭️ Skipped	39
Success Rate	88.1%
Effective Tokens	26,540,696
Estimated Cost	$15.73
Total Turns	789
Missing Tools	1
MCP Failures	0

---

❌ Failed Workflows (5)

Workflow	Run ID	Event	Tokens	Turns	Failing Job
Q	§24311376749	discussion_comment	10,104,247	103	agent
Daily Documentation Updater	§24315613017	schedule	1,389,191	162	safe_outputs
Test Quality Sentinel	§24310778700	pull_request	908,512	17	agent
Design Decision Gate 🏗️	§24310778766	pull_request	195,525	24	agent
CI Cleaner	§24313442923	schedule	155,667	21	agent

Notable: The Q workflow consumed 10.1M tokens in 103 turns on a single discussion_comment trigger — this is an outlier and likely indicates runaway exploration or an unusually complex query. The Daily Documentation Updater failed at the safe_outputs stage despite 162 turns, suggesting it produced output but the write step failed.

---

⚠️ Missing Tools (1)

• Daily DIFC Integrity-Filtered Events Analyzer: Could not access agenticworkflows audit — permission denied when verifying repository access. This tool is legitimately needed for its DIFC gateway analysis task.

---

📊 Top Token Consumers

Workflow	Runs	Effective Tokens	Avg/Run
Q	5	13,315,318	2,663,063
Test Quality Sentinel	7	3,464,057	494,865
Daily Documentation Updater	1	1,389,191	1,389,191
Daily Copilot Token Usage Audit	1	1,292,629	1,292,629
Copilot Token Usage Optimizer	1	1,220,860	1,220,860

Q accounts for 50% of all token usage across 5 runs, including one catastrophic 10M-token run. This warrants investigation.

---

🔍 Agentic Assessment Patterns (100 assessments across 85 runs)

Assessment Kind	Count
overkill_for_agentic	39
partially_reducible	29
resource_heavy_for_domain	19
model_downgrade_available	7
poor_agentic_control	6

View high-severity assessments (15 runs flagged)

Workflows flagged as resource_heavy_for_domain at HIGH severity:

• CI Cleaner — General Automation, heavy profile
• Q (×2) — Issue Response, heavy profile
• Auto-Triage Issues — Triage, heavy profile
• Sergo - Serena Go Expert — General Automation, heavy profile
• Daily Copilot Token Usage Audit — Research, heavy profile
• GitHub API Consumption Report Agent (×2) — Research, heavy profile
• Step Name Alignment — General Automation, heavy profile
• Copilot Agent Prompt Clustering Analysis — Research, heavy profile
• Design Decision Gate 🏗️ — General Automation, heavy profile
• Test Quality Sentinel — General Automation, heavy profile
• Contribution Check — General Automation, heavy profile
• Daily Documentation Updater — Repo Maintenance, heavy profile
• Static Analysis Report — Research, heavy profile

Model downgrade recommendations (7 workflows)

These workflows may not need frontier models:

• Q (×2) — Issue Response domain → consider claude-haiku-4-5 or gpt-4.1-mini
• Auto-Triage Issues — Triage domain
• PR Triage Agent — Triage domain
• Daily Workflow Updater — Repo Maintenance domain
• Daily Documentation Updater — Repo Maintenance domain
• Design Decision Gate 🏗️ — Issue Response domain

Poor agentic control signals (6 workflows)

Workflows showing broad/weakly controlled exploration with zero friction:

• Q (×2) — exploratory execution, friction=0
• Auto-Triage Issues — exploratory, read_only
• Daily Workflow Updater — exploratory, read_only
• Daily Documentation Updater — exploratory, read_only
• Design Decision Gate 🏗️ — exploratory, read_only

Recommendation: tighten instructions, reduce tool breadth, delay write actions until stronger evidence.

---

💡 Recommendations

1. Investigate Q workflow — A single discussion_comment triggered 103 turns / 10M tokens. Add turn limits or complexity guards to prevent runaway exploration.
2. Reduce model tier for triage/maintenance workflows — Auto-Triage Issues, Daily Workflow Updater, Daily Documentation Updater, PR Triage Agent, and Design Decision Gate could use smaller models (haiku/mini) to reduce cost.
3. 39 overkill_for_agentic flags — Many slash-command workflows (Issue Monster, Resource Summarizer, /cloclo, Plan Command, Archie, Scout) are running agentic inference for tasks that could be handled deterministically. Consider converting to pre-agent steps.
4. Daily Documentation Updater safe_outputs failure — 162 turns of work lost due to safe_outputs stage failing. Investigate the write failure to prevent repeated wasted compute.

---

📈 Trend Charts

Charts uploaded as workflow run artifacts.

Workflow Health (Apr 11–12): Success count tripled day-over-day (14→37) as workflow triggers became more active. The failure count also rose (1→5) but is proportionally in line with the volume increase. Skipped runs stayed roughly constant at ~40.

Token & Cost (Apr 11–12): Effective token usage increased 3× (8.5M → 26.5M) and cost nearly doubled ($8.44 → $15.73), primarily driven by the Q workflow and several heavy research/audit workflows running for the first time today.

---

References:

• §24311376749 — Q failure (10M token outlier)
• §24315613017 — Daily Documentation Updater failure
• §24316296373 — This audit run (charts in artifacts)

Generated by Agentic Workflow Audit Agent · ● 401.8K · ◷

• expires on Apr 13, 2026, 9:31 PM UTC

[github-actions[bot]]

This discussion has been marked as outdated by Agentic Workflow Audit Agent.

A newer discussion is available at Discussion #26090.