Gracefully handling tool failures

[geodavic]

I noticed that if a tool fails more than the configured max_retries, the run fails (raises an exception).

How can I gracefully allow these failures? Say a tool should fail and the agent needs to find a different approach. The main use case I'm dealing with is tools that make API calls that result in a 400-type error and raise it. I don't want the run to end just because a resource wasn't found. How can I configure this?

My first solution was to just wrap the tool in a try/except clause, and return the error string if an exception gets raised. But that feels like the wrong approach (especially since I'm dynamically generating my tools, and monkeypatching the return types to be Union[str, ...] everywhere would be messy)

[Suiranoil]

I solved this by using a custom wrapper toolset:

class InfallibleToolset(WrapperToolset[AgentDepsT]):
    """
    Catches all exceptions in tool calls and returns them as a string
    """

    async def call_tool(
        self, name: str, tool_args: dict[str, Any], ctx: RunContext[AgentDepsT], tool: ToolsetTool[AgentDepsT]
    ) -> Any:
        try:
            return await super().call_tool(name, tool_args, ctx, tool)
        except Exception as e:
            return str(e)

[DouweM]

@geodavic We're discussing a very similar issue and possible solutions in #2586. Assuming that covers it, I'll close this as a duplicate, but let me know if not and I can reopen.