Production agent pipelines need to split human rejection from execution failure

mark_automates · reddit · 2026-07-27

The author argues that in approval-gated agent workflows, a human rejection and an execution failure should be treated as different failure modes.

In a newsletter pipeline, the first workflow gathers and cleans data from five sources, then a second workflow drafts a newsletter and sends it for email approval.
If the agent times out or produces bad output, the system routes it through normal error handling.
If a human rejects the draft, that is not a broken run: the agent completed the task, but the reviewer wanted changes.
The rejected reply is parsed for feedback and passed to a separate agent for a real second attempt.
Rejections are logged separately from technical failures so operators can act on them differently.

The author says this distinction matters in production because otherwise “revise this” and “something is broken” collapse into the same queue.

More from coding & agent