[audit-workflows] Agentic Workflow Audit — 2026-10-23

[github-actions[bot]]

Summary

Across 465 unique run records collected in five pages, the adjusted success rate was 87.5% (87.1% raw), excluding two intentional-failure test runs. No missing tools were reported. The clearest shared issues are 24 failures in the Handle agent failure step, eight failures restoring daily AIC observations, recurring AI Moderator no-job runs, and incomplete log coverage for the 30-day request.

24-hour audit

• 465 unique run records, timestamped 2026-10-22 21:31 UTC through 2026-10-23 21:19 UTC; collection remains partial around page and episode boundaries.
• 442 executed runs, 385 successes, and two intentional-failure runs. The adjusted success rate is 385/440 = 87.5%; the raw rate is 385/442 = 87.1%.
• 81 reported errors: 30 driver-exit failures and 18 agent-logic failures. Missing tools: 0; warnings: 0.
• Engine counts from summary.engine_counts: copilot 176, pi 132, codex 71, claude 52, aider 3, goose 2, opencode 2, and one each for agy, crush, cursor, deepseek-harness, gemini, kiro, and pydantic-ai.
• The 30-day request returned only 100 runs spanning about eight hours. A request for 1,000 runs timed out at the five-minute limit. The charts therefore show only two observed days (465 unique runs), not a representative 30-day trend.

Findings and recommendations

1. Shared failure-handler failures: Handle agent failure failed in 24 workflows, including driver-exit cases. Fix the shared handler and trace the upstream driver failures separately so the aftermath step does not hide root causes.
2. AIC observation restore failures: Restore daily AIC scan observations failed in eight workflows. Investigate the common activation/artifact restore path and add coverage for the affected workflows.
3. AI Moderator no-job runs recur: four no-job records appear in the sample, matching the prior audit hotspot. Check whether the workflow triggers and event filters match the intended runs.
4. Token outliers: the largest sampled run used 3,795,372 tokens (Daily AWF Spec Compiler Surfacing Review); PureLock used 3,483,326. Review their token budgets and whether the work can be narrowed.
5. Log pagination: the count cap prevents dependable 30-day rollups. Repair or page the collector so the report can cover the requested window and calculate a seven-day moving average.

Workflow health chart

The two observed dates show 92.3% adjusted success on Oct 22 (60/65) and 86.7% on Oct 23 (325/375). Because 28 days are absent from the sample, this is not evidence of a month-long decline.

Token usage chart

The sample contains 8.76M tokens on Oct 22 and 50.72M on Oct 23, alongside 65 and 400 run records respectively; the difference mainly reflects sample volume and does not establish higher per-run usage. A seven-day moving average is unavailable with only two observed days.

References:

• §39864188905 — largest token outlier
• §39936302841 — daily AIC restore cluster
• §39971310352 — work-queue smoke failures

Generated by 🔍 Agentic Workflow Audit Agent · codex · gpt60 · ~7 AIC · ⌖ ~1 AIC · ⊞ 16.2K · ◷

• expires on Oct 24, 2026, 1:45 PM UTC-08:00

[github-actions[bot]]

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

A newer discussion is available at Discussion #71712.