Is your feature request related to a problem? Please describe.
As developers build autonomous loops (e.g., self-healing code agents), there is a significant risk of runaway costs due to infinite loops or hallucinations. Currently, limiting max_iterations is a proxy, but it doesn't account for token density or model cost variance.

Describe the solution you'd like
A first-class CircuitBreaker policy or middleware that can be attached to an Agent or Workflow.

• Configurable Limits: Set hard caps on TotalTokens or EstimatedCost per execution/session.
• Stateful Tracking: Accumulate usage across multiple turns in a workflow.
• Fail-Safe: If the limit is breached, the agent should halt immediately and return a safe error state (e.g., CostLimitExceededException), preventing further LLM calls.

Describe alternatives you've considered

• Manually tracking usage in every step of the workflow (error-prone and verbose).
• Setting strict max_iterations (imprecise; a single iteration with 128k context is different from 10 iterations of 1k).

Additional context
This feature is critical for "set and forget" autonomous scenarios in enterprise environments.