mukunda1729/agent-budget-violations

---
license: mit
language:
- en
tags:
- agents
- llm
- observability
- cost
- testing
- budgets
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data.jsonl
---

# agent-budget-violations

15 synthetic agent runs annotated with their **budget** (cost / tool-call / wall-time caps), **actual usage**, **violation types**, and a one-line **root cause + fix**. Built as fixtures for budget-enforcement tests, alerting heuristics, and observability dashboards.

5 of the 15 are clean (no violations) so you can test the "no false positive" path.

## Violation breakdown

| Violation type | Count |
|---|---|
| `cost` | 4 |
| `tool_calls` | 6 |
| `wall_time` | 4 |
| **None (clean)** | 5 |

(Some runs violate multiple budgets — totals don't sum.)

## Schema

```jsonc
{
"id": "string",
"agent": "string",
"budget": {
"max_tool_calls": 10,
"max_cost_usd": 1.00,
"max_wall_seconds": 60
},
"actual": {
"tool_calls": 47,
"cost_usd": 4.32,
"wall_seconds": 312
},
"violation_types": ["tool_calls", "cost", "wall_time"],
"root_cause": "string | null",
"fix": "string | null"
}
```

## Common root causes covered

- Infinite loops on tool errors
- Slow third-party APIs
- Model fallback to expensive tier
- Off-by-one budget checks
- Recursive task misinterpretation
- LLM provider rate limits
- Clarifying-question loops
- Pagination explosion

## Quickstart

```python
from datasets import load_dataset
ds = load_dataset("mukunda1729/agent-budget-violations", split="train")
multi_violators = [r for r in ds if len(r["violation_types"]) >= 2]
print(f"{len(multi_violators)} multi-budget violations")
```

## Related

- [The Agent Reliability Stack](https://mukundakatta.github.io/agent-stack/)

## License

MIT.