Error Recovery
Inspect the actual tool result and runtime status before retrying. Some workflow Activities retry under their configured Temporal policy; a tool failure may instead become an error message the agent can correct. This skill does not replace either policy with a fixed retry count or backoff schedule.
Decide what happened
- Invalid input: compare arguments with the exposed schema and correct the actual field or value.
- Missing/expired authorization: request the supported connection or owner action; do not invent credentials.
- Transient read failure: retry only when the runtime/tool permits it, respecting retry-after guidance and avoiding duplicate retries already in progress.
- Rejected mutation: correct the cause before making another authorized attempt.
- Uncertain send/write: verify the remote outcome or use the owner's reconciliation flow. A timeout is not proof that nothing happened; never blindly repeat it.
- Cancelled/reset work: stop acting under its old task identity. Cooperative Stop may still be draining admitted work.
Alternatives and partial work
Choose another connected tool only if it preserves the requested meaning and access rules. Explain stale data or unavailable sections, and distinguish a partial deliverable from completed work. Preserve useful evidence and error identities without exposing credentials or private payloads.
For managed teams, use Task Manager's revision-aware review/retry/reassignment operations instead of spawning duplicate direct delegations. Read the current task before changing it. Direct delegation, where exposed by a legacy graph, is depth-bounded and must follow its returned task identity.
Do not implement network retry loops inside Python Executor: it provides no network client, disallows imports, and its synchronous code blocks the worker. Use connected network tools and their runtime policies. Process Manager can run a fuller program when explicitly needed and permitted, but that program's side effects and cleanup remain the caller's responsibility.
