Error Handling in Goose
Error handling is a key performance-driving part of Goose. There are many ways that the non-determinism in the LLM can introduce an error that it can in turn recover from. In a typical Goose session, it's expected for there to be several agent errors that the model can see directly and correct, perhaps entirely behind the scenes.
Traditional Errors
While the agent is operating, there can be intermittent issues in the network, availability of the foundational model, etc. These are raised as errors in the agent API to the caller, who can decide how to handle that. We generally handle these with anyhow::Error.
Agent Errors
There are several types of errors where everything is working correctly, but the model generations themselves are somehow causing errors. Things like generating an unknown tool name, incorrect parameters, or a well formed tool call that results in an error in the tool itself. All of these can be surfaced to the LLM to have it attempt to recover.
The error messages are in some ways prompting - they give instructions to the LLM on how it might go about recovering. We handle these with thiserror::Error and carefully maintain a collection.
To cover all these cases, both ToolUse
and ToolResult
are typically passed through the API as part of a
Result<T, AgentError>
. An error in a ToolUse
will immediately become an error in a ToolResult
and
passed back to the LLM. A valid ToolUse
might still end up in an error ToolResult
, which is also passed
back to the LLM.
The providers then handle translating the agent errors into the various API specs as valid messages.