The Error That Never Reached the Agent

An AI Agent workflow was built to answer helpdesk queries using a set of connected tools, including a ticket management integration. The agent's system prompt described explicit error-handling logic: if a tool call fails, try looking up the correct parameters first, then retry. In practice, whenever the ticket management tool returned an error response (for example, a 400-level error from an invalid assignee) the entire workflow terminated immediately. The error was not passed back to the agent as a tool observation. The agent never saw the error, never had a chance to apply its recovery logic, and never produced an answer.

A deliberate hard-stop policy (tool errors should always abort) considered and ruled out. The agent had explicit recovery instructions for exactly this scenario, and the framework provides no mechanism to distinguish 'user intended this as a hard stop' from 'user intended the agent to handle this.' The termination was not a policy; it was an unintended side effect of how errors were routed.

A model failure to apply its own recovery instructions, considered and ruled out. The agent never received the error signal to act on. Distinguishing whether the model would have recovered correctly is impossible because the framework ensured it was never given the chance.

Apraxia, most consistent with the evidence: the agent had been given correct recovery logic and the right tools to apply it, but the framework's error routing made the tools effectively unusable, a specific, known type of failure (a 400-level error) could not be handled because the mechanism for communicating it to the agent was absent.

An agent's system-prompt logic can only work on what the agent can see. When an error terminates the workflow before it reaches the model, any error-handling guidance in the system prompt becomes dead letter, not because the agent won't follow it, but because the information never arrives. Framework-level choices about what counts as a 'workflow error' versus what should be returned as a 'tool observation' have direct consequences for what kinds of failures an agent can actually recover from.

Return tool errors to the agent as observations rather than terminating the workflow.

Framework error handling that surfaces tool failures to the model; testing of error paths.