Wait to run the Agent first step until Input guardrails is complete

[rach]

The guardrail may trigger a Tripewire but the agent still run

In v0.0.7, the code runs the input guardrail task and the agent's first step together asynchronously.

If the guardrail is in place to avoid a side effect from the Agent run, the agent may still do something unexpected.

Let's say that you have a tripwire guardrail: "Don't allow to delete a file", and the agent is an MCP file system agent

[rm-openai]

the current behavior is the intended/documented behavior. If you want to run the guardrail before the agent runs, you can just do that via python code:

triggered = await do_something()
if not triggered:
  await Runner.run(...)

[rach]

For my own understanding. What was the rational to have the agent start before the guardrail is done?
I can work around it, now that I know the behavior.

[rm-openai]

@rach for latency. If the guardrail doesn't fail, then you will have made a bunch of progress on the actual agent. RUnning (guardrail, agent) in parallel is much faster than serially