Vault AI Agent Limits
When executing agent actions programmatically, Vault enforces the following limits.
Per-action limits apply to a single agent action execution:
- Maximum 10 LLM calls per action.
- Maximum 20 steps per action (each LLM call counts as one step; each tool call counts as one step).
- Maximum 5 parallel tool calls per action.
Root-action limits apply across the full execution tree initiated by the outermost agent action, including any agent actions invoked through tool calls:
- Maximum 20 LLM calls total across the root action tree.
- Maximum 40 steps total across the root action tree.
When a limit is exceeded at runtime, Vault returns an appropriate error subtype.