We need to provide a mechanism so that a human user can give direct feedback to the agent during its execution. This will allow for real-time intervention, guidance, or corrections, enabling human-in-the-loop (HITL) workflows.
Requested Feature
- Implement a system for human users to interact with the agent during its tasks using the Strands Agents framework.
- Feedback should be actionable and able to influence ongoing agent actions.
- Human should be able to:
- Provide suggestions
- Provide corrections
- Pause/resume agent execution
- Feedback is unrestricted and should not be limited to any particular stage of execution.
- No additional security or logging requirements.
Additional Context Example
Cyber-AutoAgent is deployed within a corporate network to perform automated penetration testing. The agent is configured to scan for vulnerabilities, exploit weaknesses, and report findings in real time. During execution, the agent attempts to exploit a recently discovered vulnerability in an internal web application.
As the agent initiates its exploit routine, a human pentester observes the agent’s command output and notices that the agent is about to run a potentially destructive payload. Using the human-in-the-loop feedback system, the pentester pauses the agent’s execution and reviews the planned actions.
The pentester provides corrections, suggesting a safer, read-only exploit command to avoid disrupting production services. The agent resumes execution, incorporating the human’s suggestion. Later, as the agent identifies an unusual network configuration, the pentester adds contextual notes and suggestions for additional enumeration steps, improving the thoroughness of the test.
We need to provide a mechanism so that a human user can give direct feedback to the agent during its execution. This will allow for real-time intervention, guidance, or corrections, enabling human-in-the-loop (HITL) workflows.
Requested Feature
Additional Context Example
Cyber-AutoAgent is deployed within a corporate network to perform automated penetration testing. The agent is configured to scan for vulnerabilities, exploit weaknesses, and report findings in real time. During execution, the agent attempts to exploit a recently discovered vulnerability in an internal web application.
As the agent initiates its exploit routine, a human pentester observes the agent’s command output and notices that the agent is about to run a potentially destructive payload. Using the human-in-the-loop feedback system, the pentester pauses the agent’s execution and reviews the planned actions.
The pentester provides corrections, suggesting a safer, read-only exploit command to avoid disrupting production services. The agent resumes execution, incorporating the human’s suggestion. Later, as the agent identifies an unusual network configuration, the pentester adds contextual notes and suggestions for additional enumeration steps, improving the thoroughness of the test.