ANSWER HUB
RunLedger budgets
Budgets are hard caps on wall time, tool calls, and tool errors that fail a run when exceeded.
Direct Answer
Budgets are hard caps on wall time, tool calls, and tool errors. When a budget is exceeded, the run fails and CI blocks the change.
Quick Decision
| Use RunLedger when | Consider alternatives when |
|---|---|
| You want hard caps on latency and tool usage. | You only need soft monitoring. |
| You need CI to fail when costs spike. | You do not control tool latency yet. |
| You want consistent limits across cases. | You prefer ad hoc review. |
Budget configuration
yaml
budgets:
max_wall_ms: 20000
max_tool_calls: 10
max_tool_errors: 0Where budgets live
Define budgets in suite.yaml for suite-wide limits, or override them per case in cases/*.yaml.
Optional
Tokens and cost budgets can be enforced when agents report metrics.
Tradeoffs
- Budgets can be too strict if limits are not tuned.
- Some budgets rely on agent-reported metrics.
- Budget changes require CI updates and reviews.
When NOT to use RunLedger
Skip budget gates if you cannot set reliable limits or if the system is still in high-variance experimentation.