Error handling

Errors that arise during the execution of an autonomous AI agent system are categorized in two different classes: AppExceptions and AppErrors. The main difference between the two is the need of handling them, so we use this language so that developers know what must be handled by the application and what not.

AppExceptions

AppExceptions are runtime errors that arise from faulty or malicious agent-to-agent interactions and need to be caught and handled. Examples of this are transaction signature errors (SignatureNotValidError), unexpected content being sent by a faulty agent instance (TransactionNotValidError) or problems that happen while committing a new block to the blockchain (AddBlockError).

In this scenario, an error code different from 0 will be returned to the Tendermint node while the agent instance tries to recover from the error. If this is not possible, and the agent instance is not able to progress with the round, the FAIL flag will be set by calling set_fail() and the agent instance will transition into a WaitBehaviour state. In the case the round cannot be finished without this agent instance (for example, due to not meeting the minimum quorum requirements), all agent instances will try to agree to restart the round from the beginning.

AppErrors

AppErrors errors are basically inconsistencies in the application logic and are not handled. These are usually problems within the ABCI app implementation, that should not be caught, for instance missing None or initialization checks that lead to an application crash so the developers can trace it and solve it.