Retry & Failover
Automatic fallback on error, with no context loss or visible interruption.
AI systems fail. A model can be temporarily unavailable, an API can hit its rate limit, a tool can return an unexpected error.
The retry and failover system automatically handles these situations. When an action fails, the system evaluates the error's nature and applies the appropriate strategy: retry with exponential backoff, failover to an alternative model, or graceful degradation.
All of this happens transparently. The user doesn't see the retries or the failovers. They receive their result, perhaps with a slightly longer delay, but with no interruption.
Detection
Every call (model, tool, API) is wrapped in a circuit breaker that detects errors and timeouts.
Classification
The error is classified: temporary (rate limit, network timeout), permanent (endpoint removed, access denied), or unknown.
Retry strategy
For temporary errors: retry with exponential backoff (1s, 2s, 4s...) and configurable attempt count.
Failover
For permanent errors: failover to the configured alternative provider/model. Context is transferred.
Graceful degradation
If all alternatives fail: the agent clearly informs the user and suggests manual alternatives.
Rate limiting
The primary model hits its rate limit. The system automatically switches to the secondary model. The user notices nothing.
Provider outage
An API provider is under maintenance. Failover activates the alternative provider. In-flight requests are replayed without context loss.
Failing tool
An external tool returns intermittent errors. The circuit breaker temporarily disables it and the agent uses an alternative method.
featurePages.retryFailover.config
- Active by default on all EasyClaw v2 agents
- Fallback providers and models are configurable
- Compatible with all models and tools supported by EasyClaw