Self-Correction Loop for Multi-Step Tool Chains

You are a Reliability Engineer for Agentic Workflows.

Context

An agent is executing a multi-step plan to achieve {{objective}}. It must monitor its progress against the {{tolerance_threshold}} and utilize {{recovery_tools}} if the plan fails or the output quality drops.

Task

1. Draft the 'Pre-Execution Validation' step where the agent predicts the expected output of the next tool call.
2. Define the 'Deviation Detection' logic: how the agent compares actual output vs. predicted output.
3. Design the 'Pivot Logic'—a set of rules for when to continue, when to retry with different parameters, and when to re-plan entirely.
4. Specify the 'State Rollback' procedure using recovery tools to clear corrupted context or variables.
5. Create a 'Final Verification Gate' that confirms the objective is met before terminating.

Constraints

• MUST NOT allow infinite loops; define a hard retry limit per step.
• MUST require the agent to document why it is pivoting in its internal thought trace.
• MUST use the recovery tools only as a last resort before human escalation.

Output format

• Monitor-Evaluate-Act (MEA) Loop Definition
• Validation Heuristics Table
• Re-planning Decision Tree
• Log Format for Self-Correction Events

Quality bar

• Is the pivot logic unambiguous?
• Does the rollback procedure prevent state contamination?
• Are the retry limits clearly defined?