Result: A self-correcting loop.
I used this to fix my Code Reviewer agent. It found tool failures I missed and patched its own context-blindness.

Stop doing "vibes-based" prompting. Start debugging.

Step 3: Patching (`edit_file`) 🩹
The agent updates its own prompt file.
- Hardening guardrails
- Fixing schema hallucinations
- Injecting missing context

Step 2: Source Review (`Read`) 📂
The agent inspects its own "source code"—the markdown files defining its persona.
It finds the gap between what you *wanted* and what you *wrote*.

Step 1: Log Analysis (`read_thread`) 🕵️‍♂️
Don't just retry. Feed the failed chat history back to the agent. Ask:
"Why did you make that decision?"
"Where did the logic break?"