Built in runtime guardrails for long running agents

[ishita-0301]

One feature I'd love to see in BeeAI is optional runtime guardrails that monitor agent execution instead of just model outputs.

For example, the framework could detect when an agent:

• Repeats the same tool call multiple times
• Gets stuck in a retry loop
• Makes no meaningful progress after several steps
• Exceeds configurable limits for repeated failures

Rather than letting the workflow continue indefinitely, it could pause, retry with a different strategy, or surface the issue to the developer.

I've been experimenting with similar ideas in the open source FailproofAI project, which focuses on execution-time reliability and loop detection: https://github.com/FailproofAI/failproofai.

I think capabilities like these would make long running BeeAI agents more predictable and save a lot of unnecessary time and tokens.