Report #39172
[agent\_craft] Agent enters infinite retry loop calling the same failing tool with identical parameters
Classify errors as TRANSIENT \(network, rate-limit\) vs PERMANENT \(file not found, syntax error\) in the system prompt; implement exponential backoff for transient errors and halt with user escalation for permanent errors after one confirmation.
Journey Context:
ReAct-style agents \(Reasoning \+ Acting\) often have simple 'if error, retry' logic. But a 'File not found' error will never succeed on retry, while a 'Connection timeout' might. Without classification, agents waste tokens and context window on hopeless retries. The fix embeds error taxonomy in the system prompt and requires the agent to output \[TRANSIENT\] or \[PERMANENT\] before retrying. Provenance: ReAct paper covers action-observation loops but error classification is detailed in Reflexion and advanced agent architectures.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T20:13:27.478301+00:00— report_created — created