Structured error responses

• The MCP isError flag marks a tool result as a failure.
• Four categories: transient (timeouts, service unavailable), validation (bad input), business (policy violation), permission.
• Uniform generic errors prevent the agent from choosing the right recovery.
• Retryable versus non-retryable metadata prevents wasted retry attempts.
• An access failure needs a retry decision; an empty result is a successful query with no matches.

• Return errorCategory, an isRetryable boolean, and a human-readable description.
• Mark business rule violations as not retryable with a customer-friendly explanation.
• Recover from transient failures locally inside a subagent; propagate only unresolved errors, with partial results and what was attempted.
• Distinguish access failures from valid empty results.

• "The model will work out what went wrong from any message." It needs structure to decide.
• "Every error should be retried a fixed number of times." Validation and business errors never succeed on retry.
• "An empty list and a timeout can be reported the same way." They call for different decisions.

Generic failure text, errors silently returned as empty success, and whole workflows terminated on a single transient failure.

• 5.3Error propagation across agents
• 1.1Agentic loops
• 1.2Coordinator-subagent orchestration