Agent Self-Correction

Agent Self-Correction

[Context: Errors and roadblocks are inevitable. An efficient agent does not stop at the first sign of failure; it analyzes the error, adjusts its approach, and retries. This skill defines the cognitive loop for autonomous self-correction.]

1. Core Principles

• Analyze First: Never retry the exact same action without understanding why it failed.
• Fail Fast, recover Faster: Identify failure modes immediately (tech stack mismatch, syntax error, hallucination).
• Graceful Escalation: Attempt 2-3 autonomous fixes before escalating to the user.
• State Preservation: Don't destroy partial progress when fixing a specific bug.

2. Step-by-Step Implementation

Phase 1: Failure Detection & Analysis

When a tool returns an error or a check fails:

1. Stop: Do not proceed to the next step of the plan.
2. Read: Analyze the error message thoroughly.
   • Is it a syntax error? -> Fix code.
   • Is it a logic error? -> Rethink algorithm.
   • Is it a tool misuse? -> Check tool definition.
   • Is it environmental? -> Check paths/permissions.
3. Hypothesize: Formulate a hypothesis for the fix.

Phase 2: autonomous Recovery Loop

1. Attempt 1 (Quick Fix): Address the obvious error (e.g., install missing package, fix typo).
2. Verify: Run the verification command (e.g., npm run build or pytest).
3. Attempt 2 (Alternative Approach): If Attempt 1 fails, try a different distinct strategy (e.g., switch library, rewrite logic).
4. Attempt 3 (Simplify): Reduce complexity or scope to isolate the issue.

Phase 3: Escalation

If Attempt 3 fails:

1. Gather Context: Collect error logs and the 3 failed attempts.
2. Propose solution: Present the user with the situation and a proposed manual intervention or request for clarification.

3. Templates & Examples

Internal Monologue for Debugging

ERROR DETECTED: [Error Message]
ANALYSIS: The error implies [Cause].
HYPOTHESIS: If I change [X] to [Y], it should resolve.
ACTION: Applying fix [X->Y].

Self-Correction Checklist

• Did I read the entire error message?
• Have I checked if the file path is correct?
• Did I verify the library version compatibility?
• Is this a hallucinated import?

4. Common Pitfalls

• Don't: Blindly loop the same command hoping it works eventually.
• Don't: Delete the user's codebase to "start over" without permission.
• Do: Use common_sense to judge if an error is a simple typo or a fundamental design flaw.