[agent-efficiency] Agent runs fail after pick-three sub-agent fanout

[github-actions]

Run summary

Lookback window: 2026-06-16T17:38:18Z to 2026-06-23T17:38:18Z. These counts are for agentic workflow paths containing trigger- or gh-aw-. Pass/fail rates are calculated over success + failure runs only; startup failures and non-terminal/approval statuses are shown separately.

Repository	Total runs	Success	Failure	Startup failure	Other statuses	Pass rate	Fail rate
elastic/ai-github-actions	1,115	84	109	36	886	43.5%	56.5%

Other statuses for this repository were action_required 777, skipped 106, cancelled 2, and in_progress 1. No downstream repository rows are included: downstream discovery via GitHub code search was unavailable from this run (search_code returned 429 and unauthenticated public REST returned 401), so no downstream repositories were reliably discovered.

Findings

1. Pick-three workflows fan out to sub-agents and then fail on provider auth or AI credit limits

Affected downloaded failed runs: 4 of the 9 logs where the agent process started: trigger-docs-patrol, trigger-text-auditor, trigger-framework-best-practices, and trigger-bug-hunter.

Evidence:

• Docs Patrol run https://github.com/elastic/ai-github-actions/actions/runs/28035785847 launched three General-purpose(gpt-5.3-codex) sub-agents at /tmp/gh-aw/logs/28035785847/4_run _ agent.txt:1684-1691, then each read returned Authentication failed with provider at (172.30.0.30/redacted) (HTTP 403) at lines 1699, 1705, and 1711. The run ended with Tokens ↑ 1.8m (1.5m cached) • ↓ 21.5k at line 1717 and failureClass=authentication_failed at line 1720.
• Text Auditor run https://github.com/elastic/ai-github-actions/actions/runs/28031348044 launched three General-purpose(gpt-5.3-codex) sub-agents at /tmp/gh-aw/logs/28031348044/4_run _ agent.txt:1595-1602, then logged provider 403s at lines 1613, 1619, 1625, and 1631. The run ended with Tokens ↑ 1.8m (1.3m cached) • ↓ 24.5k at line 1637 and failureClass=authentication_failed at line 1640.
• Framework Best Practices run https://github.com/elastic/ai-github-actions/actions/runs/28031380597 logged repeated provider 403s at /tmp/gh-aw/logs/28031380597/4_run _ agent.txt:1626-1650 and again at lines 1669, 1672, and 1682. It ended with Tokens ↑ 1.9m (1.0m cached) • ↓ 26.3k at line 1688 and failureClass=authentication_failed at line 1691.
• Bug Hunter run https://github.com/elastic/ai-github-actions/actions/runs/28024841420 launched three General-purpose(gpt-5.3-codex) sub-agents at /tmp/gh-aw/logs/28024841420/4_run _ agent.txt:1754-1761, then hit CAPIError: 429 Maximum AI credits exceeded (1057.625500 / 1000) at lines 1775 and 1777. The run ended with Tokens ↑ 2.0m (1.4m cached) • ↓ 20.9k at line 1782 and AI credits budget exceeded — not retrying at line 1786.

Root cause: The shared prompt fragments require this fanout without a budget or model-availability guard. .github/workflows/gh-aw-fragments/pick-three-keep-many.md:3-15 says to spawn multiple independent sub-agents, make each prompt fully self-contained, and wait for all sub-agents. .github/workflows/gh-aw-fragments/pick-three-keep-one.md:3-15 similarly requires exactly three sub-agents and explicitly says a 10,000-token prompt is preferable to a short one. These fragments are included by the affected workflow sources at .github/workflows/gh-aw-docs-patrol.md:15, .github/workflows/gh-aw-text-auditor.md:14, .github/workflows/gh-aw-framework-best-practices.md:14, and .github/workflows/gh-aw-bug-hunter.md:15.

2. Broad GitHub MCP issue searches overflowed in failed runs

Affected downloaded failed runs: 2 of the 9 logs where the agent process started: trigger-framework-best-practices and trigger-bug-hunter.

Evidence:

• Framework Best Practices run https://github.com/elastic/ai-github-actions/actions/runs/28031380597 called search_issues (MCP: github) with an open-issue title query at /tmp/gh-aw/logs/28031380597/4_run _ agent.txt:1594, and the tool response was Output too large to read at once (57.8 KB). Saved to: /tmp/1782222986168... at line 1595.
• Bug Hunter run https://github.com/elastic/ai-github-actions/actions/runs/28024841420 called search_issues (MCP: github) · bug hunter OR bug-hunter at /tmp/gh-aw/logs/28024841420/4_run _ agent.txt:1730, and the tool response was Output too large to read at once (59.7 KB). Saved to: /tmp/1782216509451... at line 1731.

Root cause: The agent used broad MCP issue searches for duplicate checks in workflows whose prompts also require sub-agent fanout. The repository prompt includes MCP pagination guidance, but the observed queries did not constrain result size enough to avoid large dumped responses before the later sub-agent/provider failures.

Duplicate check

Open issue searches for Authentication failed with provider + sub-agent, Maximum AI credits exceeded + sub-agent, Pick Three + sub-agents + AI credits, and pick-three-keep + Authentication failed found no matching open tracking issue. The separate UV_PATH startup failure in trigger-code-duplication-detector and trigger-code-complexity-detector is already covered by open draft PR #1385, so it is not reported here.

Generated by Internal: Agent Efficiency · 936.5 AIC · ⌖ 12.1 AIC · ⊞ 24.5K · ◷

• expires on Jun 30, 2026, 5:54 PM UTC