[audit-workflows] Agentic Workflow Audit — 2026-05-04

[github-actions[bot]]

Overview

Daily audit of all agentic workflow runs from the last 24 hours (2026-05-04). 50 runs analyzed across copilot and claude engines. Zero failures detected — all completed runs succeeded. The primary concern today is resource efficiency: 5 workflows were flagged as resource-heavy for their task domain, and one workflow experienced significant network friction (47% firewall block rate).

Run Summary

Metric	Value
Total Runs	50
✅ Success	12
⏭️ Skipped	35
❌ Failed	0
🔄 In Progress	3
Success Rate (of completed)	100%
Total Tokens	5,870,876
Effective Tokens	4,169,282
Total Cost	$2.26
Total Action Minutes	119 min
Total Turns	177

Engine Distribution

Engine	Runs
Copilot	8
Claude	4
Unknown (skipped/queued)	38

---

Workflow Health Chart

The workflow health chart shows a clean run profile for 2026-05-04 with zero failures. Activity is concentrated in two time windows (19:00–20:00 UTC and 21:00–22:00 UTC), with the bulk of skipped runs representing PR-triggered workflows that had no qualifying events. The 100% success rate across all completed runs indicates healthy workflow infrastructure.

---

Token & Cost Chart

Cost is concentrated in three claude-engine runs totaling $2.26 for the day. The Lockfile Statistics Analysis Agent was the most expensive single run at $0.80 with 41 turns — significantly more turns than task complexity warrants for a research/analysis workflow. Token usage peaked during the 21:00 UTC hour when the claude-engine daily workflows executed.

---

⚠️ High Priority Issues

1. Resource-Heavy Runs (5 workflows)

Five workflows were flagged as consuming a heavy execution profile for their task domain:

Workflow	Turns	Duration	Domain
Lockfile Statistics Analysis Agent	41	7m 40s	Research
Contribution Check	56	8m 2s	General Automation
Daily Project Performance Summary	30	17m 59s	General Automation
Daily Caveman Optimizer	14	8m 52s	General Automation
Daily Team Evolution Insights	16	6m 22s	General Automation

Recommendation: Move data-fetching steps to deterministic frontmatter steps (pre-agent writes to /tmp/gh-aw/agent/) to reduce agentic inference turns. About 50% of turns in each run are data-gathering that could be handled deterministically.

2. Network Friction Hotspot

The Daily Project Performance Summary Generator (Using MCP Scripts) had 36 blocked requests out of 77 total (47% block rate). This is the highest firewall block pressure observed today.

Recommendation: Audit which domains are being blocked and either whitelist legitimate domains in the workflow's firewall config or switch to alternative API endpoints.

---

Medium Priority Issues

Execution Drift — Issue Monster (3 occurrences)

The Issue Monster workflow varied significantly across runs:

• Run 25344613337: 0 turns (skipped/trivial)
• Run 25342963037: 6 turns
• Run 25340618352: 5 turns

Average: 3.7 turns, range 0–6. This suggests changing task shape or unstable prompts. Consider adding more deterministic filtering before invoking the agent.

Model Downgrade Opportunity — Issue Monster

Two Issue Monster runs (25342963037, 25340618352) were flagged as candidates for model downgrade. The issue-response domain with read-only actuation does not require a frontier model.

Recommendation: Add engine.model: claude-haiku-4-5 or gpt-4.1-mini to the Issue Monster workflow frontmatter to reduce costs.

Anomaly Events (2 detected)

Cross-run analysis detected 2 events with anomaly score > 0.6, indicating patterns unusual relative to learned templates. Both were tool_result stage events with score 0.65 — new log template discovered in a rare cluster with few prior observations. This is low-risk but worth monitoring as the pattern accumulates more data.

Overkill-for-Agentic Pattern (30 skipped runs)

Multiple skipped issue-response workflows (Scout, Q, Archie, ACE Editor, Plan Command, Mergefest, etc.) are flagged as potentially overkill for agentic automation. These all have 0 turns and read-only actuation. This is expected behavior for PR-triggered workflows that skipped due to no qualifying events, but worth reviewing if these workflows are intended to be more active.

---

Top 3 Costly Runs

Workflow	Cost	Tokens	Turns	Conclusion
Lockfile Statistics Analysis Agent	$0.80	969,260	41	✅ success
Daily Team Evolution Insights	$0.76	793,610	16	✅ success
Daily Caveman Optimizer	$0.70	503,614	14	✅ success

---

MCP Tool Usage (Last 24h)

Server	Tool	Calls
safeoutputs	assign_to_agent	6
safeoutputs	add_comment	6
safeoutputs	create_discussion	5
github	issue_read	4
safeoutputs	upload_asset	3
safeoutputs	noop	3
github	list_pull_requests	1
github	list_issues	1
mcpscripts	github_pr_query	1

---

Recommendations Summary

1. Reduce turns in Lockfile Stats & Contribution Check — Move data-gathering to deterministic pre-agent steps to cut 20-30 turns each
2. Fix Daily Project Performance firewall config — 47% block rate suggests misconfigured allowed domains
3. Downgrade Issue Monster model — Use claude-haiku-4-5 for issue-response domain to reduce cost
4. Stabilize Issue Monster prompts — Address 0–6 turn drift to improve predictability

References:

• §25343247456 — Daily Caveman Optimizer (resource heavy)
• §25342362056 — Lockfile Statistics (resource heavy, highest cost)
• §25343424019 — Daily Project Performance (network friction)

Generated by Agentic Workflow Audit Agent · ● 574.5K · ◷

• expires on May 5, 2026, 9:59 PM UTC