Description
The jsweep workflow exploded from 8 turns (baseline, 2026-04-21) to 60 turns on 2026-04-24, consuming 4.64M tokens and failing after 25 minutes. Only 1 tool type was called across all 60 turns (safeoutputs.create_pull_request), indicating the agent looped without making progress. The previous baseline run completed in ~5 min and ~8 turns.
Audit comparison:
| Metric |
Baseline (2026-04-21) |
Failing run (2026-04-24) |
| Turns |
8 |
60 (+650%) |
| Duration |
~5 min |
25.1 min |
| Token usage |
low |
4.64M |
| Result |
success |
failure |
Expected Impact
Prevents token waste of 4.64M+ per failed run. Setting max_turns: 20 would cap the cost at ~620K tokens, a ~87% reduction in runaway cases.
Suggested Agent
An agent should: (1) add max_turns: 20 to the jsweep workflow frontmatter, (2) review the prompt for ambiguous success criteria and add explicit done conditions, and (3) check the diff between the April 21 and April 24 runs to identify what changed.
Estimated Effort
Quick (< 1 hour) — frontmatter edit + prompt clarification.
Data Source
DeepReport Intelligence Briefing — April 24, 2026 (run §24896194233). Full investigation in #28218.
Generated by DeepReport - Intelligence Gathering Agent · ● 570.7K · ◷
Description
The jsweep workflow exploded from 8 turns (baseline, 2026-04-21) to 60 turns on 2026-04-24, consuming 4.64M tokens and failing after 25 minutes. Only 1 tool type was called across all 60 turns (
safeoutputs.create_pull_request), indicating the agent looped without making progress. The previous baseline run completed in ~5 min and ~8 turns.Audit comparison:
Expected Impact
Prevents token waste of 4.64M+ per failed run. Setting
max_turns: 20would cap the cost at ~620K tokens, a ~87% reduction in runaway cases.Suggested Agent
An agent should: (1) add
max_turns: 20to the jsweep workflow frontmatter, (2) review the prompt for ambiguous success criteria and add explicit done conditions, and (3) check the diff between the April 21 and April 24 runs to identify what changed.Estimated Effort
Quick (< 1 hour) — frontmatter edit + prompt clarification.
Data Source
DeepReport Intelligence Briefing — April 24, 2026 (run §24896194233). Full investigation in #28218.