Principle:HKUDS AI Trader Agentic Retry Loop
| Knowledge Sources | |
|---|---|
| Domains | Reliability, LLM_Agents |
| Last Updated | 2026-02-09 14:00 GMT |
Overview
A fault-tolerance pattern that wraps LLM invocations with exponential backoff retries to handle transient API failures.
Description
Agentic Retry Loop provides resilience against transient failures in LLM API calls. External LLM providers may return rate limit errors (429), server errors (500+), or timeout on any individual request. Rather than failing the entire trading session, this pattern retries the invocation with exponentially increasing delays.
This is essential for multi-day backtesting where a single API failure should not abort the entire simulation. The retry mechanism wraps each individual LLM invocation, separate from the multi-step reasoning loop.
Usage
Use this principle wherever LLM inference calls are made against external APIs. It is the inner retry layer (per-invocation), distinct from the outer reasoning loop (multi-step trading session).
Theoretical Basis
# Pseudocode for retry with exponential backoff
for attempt in range(1, max_retries + 1):
try:
response = await agent.ainvoke(messages)
return response
except Exception as e:
if attempt == max_retries:
raise
delay = base_delay * attempt
await sleep(delay)
Key properties:
- Exponential backoff: Delay increases linearly with attempt number (base_delay * attempt)
- Bounded retries: Maximum number of attempts prevents infinite loops
- Transparent: Callers receive the response as if no retry occurred
- Last-exception propagation: Final failure re-raises the last exception