[agent-efficiency] Current 7-day agent runs still fail after Pick Three sub-agent fanout

[github-actions]

Lookback analyzed: 2026-07-02T17:39:43Z to 2026-07-09T17:39:43Z (last 7 days).

Sources used: /tmp/gh-aw/errors.json, /tmp/gh-aw/failed_runs.json, /tmp/gh-aw/logs/<run_id>/..., GitHub MCP code search for downstream workflow users, and GitHub MCP workflow-run metadata pages. Downstream rows are metadata-only; no downstream logs were downloaded or inspected.

Per-repository run summary

Repository	Metadata coverage	Agentic runs observed	Success	Failure	Other	Pass rate
elastic/ai-github-actions	Latest GitHub MCP page plus 20 predownloaded failed runs from 2026-07-07 to 2026-07-09	24 latest-page rows; 20 failed-run corpus entries	12	10 latest-page; 20 failed corpus	1 queued, 1 in_progress	54.5% on latest-page completed runs
elastic/beats	Latest GitHub MCP page, metadata only	30	0	0	30 skipped	n/a
elastic/logstash	Latest GitHub MCP page, metadata only	19	0	0	19 skipped	n/a
elastic/integrations	Latest GitHub MCP page, metadata only	16	0	0	16 skipped	n/a
elastic/elastic-agent	Latest GitHub MCP page, metadata only	30	0	15	15 skipped	0.0% of non-skipped completed runs
elastic/elastic-ramen	Latest GitHub MCP page, metadata only	2	1	0	1 skipped	100.0% of non-skipped completed runs
elastic/infra-party	Latest GitHub MCP page; code-search result was Claude composite usage, not gh-aw lock workflow usage	0	0	0	0	n/a

Findings

1. Pick Three workflows repeatedly fail after sub-agent fanout with provider authentication errors

What happened: 11 of the 20 predownloaded failed runs show the parent agent launching general-purpose sub-agents, consuming about 1.8M-1.9M input tokens, then ending with a provider authentication failure. The affected workflows are:

Workflow	Failed runs
.github/workflows/trigger-autonomy-atomicity-analyzer.yml	3
.github/workflows/trigger-text-auditor.yml	3
.github/workflows/trigger-bug-hunter.yml	3
.github/workflows/trigger-framework-best-practices.yml	1
.github/workflows/trigger-information-architecture.yml	1

Evidence: In Autonomy Atomicity Analyzer run https://github.com/elastic/ai-github-actions/actions/runs/29035122191, the agent launched three independent general-purpose background agents at /tmp/gh-aw/logs/29035122191/run _ agent/35_Execute GitHub Copilot CLI.txt:266-273:

General-purpose(gpt-5.3-codex) Analyze shared state
  Agent started in background with agent_id: state-hotspots
General-purpose(gpt-5.3-codex) Analyze registries
  Agent started in background with agent_id: registry-hotspots
General-purpose(gpt-5.3-codex) Analyze test coupling
  Agent started in background with agent_id: test-ordering-hotspots

The same run then failed at /tmp/gh-aw/logs/29035122191/run _ agent/35_Execute GitHub Copilot CLI.txt:344-356:

Authentication failed with provider [redacted] (HTTP 403).
Check your COPILOT_PROVIDER_API_KEY or COPILOT_PROVIDER_BEARER_TOKEN.
Tokens     ↑ 1.8m (1.3m cached) • ↓ 19.5k (8.0k reasoning)
[copilot-harness] attempt 1 failed: exitCode=1 failureClass=authentication_failed ... hasOutput=true retriesRemaining=3
[copilot-harness] attempt 1: Copilot requests authentication failed through the gh-aw API proxy (HTTP 403, model=gpt-5.3-codex, stage=starting the Copilot CLI request).
— not retrying (first-attempt auth failure is non-retryable)

Bug Hunter run https://github.com/elastic/ai-github-actions/actions/runs/29016891696 shows the same pattern at /tmp/gh-aw/logs/29016891696/4_run _ agent.txt:1637-1648 and /tmp/gh-aw/logs/29016891696/4_run _ agent.txt:1676-1701: three general-purpose sub-agents launched, repeated wait messages, then HTTP 403 provider authentication failure with Tokens ↑ 1.9m (1.4m cached).

Root cause: The affected workflow prompts require Pick Three fanout through hand-authored workflow markdown and shared fragments. Examples:

.github/workflows/gh-aw-autonomy-atomicity-analyzer.md:114
Use the Pick Three, Keep One pattern ... spawn 3 general-purpose sub-agents ... Include the architecture overview, module structure, and the full criteria in each sub-agent prompt.

.github/workflows/gh-aw-bug-hunter.md:107
Use the Pick Three, Keep One pattern ... spawn 3 general-purpose sub-agents ... Include the git log output, repo conventions, and the full criteria in each sub-agent prompt.

.github/workflows/gh-aw-fragments/pick-three-keep-one.md:3-15
Parallelize your initial investigation using sub-agents ... Each prompt must be fully self-contained ... Wait for all 3 sub-agents to complete.

2. AI-credit/unknown-model conclusion flags recur in the same fanout-heavy workflows

What happened: 7 failed runs have AI credits rate-limit error: true; 5 of those also have Unknown model AI credits error: true. The conclusion logs associate those flags with the same failed parent-agent executions.

Workflow	Runs with AI-credit flag
.github/workflows/trigger-autonomy-atomicity-analyzer.yml	3
.github/workflows/trigger-text-auditor.yml	2
.github/workflows/trigger-framework-best-practices.yml	1
.github/workflows/trigger-bug-hunter.yml	1

Evidence: Autonomy Atomicity Analyzer run https://github.com/elastic/ai-github-actions/actions/runs/29035122191 records at /tmp/gh-aw/logs/29035122191/0_run _ conclusion.txt:466-475:

AI credits: 1039.951
Configured max AI credits: 1000
AI credits rate-limit error: true
Max AI credits exceeded (harness budget abort): false
Daily workflow AIC guardrail exceeded: false
Inference access error: false
MCP policy error: false
Agentic engine timeout: false
Model not supported error: false
Unknown model AI credits error: true

Framework Best Practices run https://github.com/elastic/ai-github-actions/actions/runs/29023339978 records the same conclusion flags at /tmp/gh-aw/logs/29023339978/1_run _ conclusion.txt:467-476 after an agent transcript that launched three general-purpose scans and ended with Tokens ↑ 1.9m (1.4m cached).

Root cause: The root prompt pattern is the same Pick Three sub-agent fanout above. Additional directly affected prompt entry points include .github/workflows/gh-aw-text-auditor.md:150-151, .github/workflows/gh-aw-framework-best-practices.md:121-122, and .github/workflows/gh-aw-information-architecture.md:114, each requiring three general-purpose sub-agents with self-contained prompts.

Duplicate check

Open issue #1519 tracks the previous 2026-06-29 to 2026-07-06 7-day window. This report covers the fresh 2026-07-07 to 2026-07-09 failed-run corpus within the current 7-day lookback and shows the same provider-auth/AI-credit pattern recurring in new runs.

The UV_PATH setup failure also appeared in 7 current failed runs across Code Complexity Detector and Code Duplication Detector, but it is not included as a primary finding here because open draft PR #1385 directly identifies and addresses that exact Expose uv in workspace failure mode.

Generated by Internal: Agent Efficiency · 739.2 AIC · ⌖ 16.5 AIC · ⊞ 25.3K · ◷

• expires on Jul 16, 2026, 5:45 PM UTC

[github-actions]

This issue was automatically closed because it expired on 2026-07-16T17:45:45.961Z.

Closed by Workflow