How Small Teams Should Assign Responsibilities to AI Agents

The most common pitfall for AI teams isn’t a lack of agents, but rather a situation where every agent seems to be “somewhat responsible.” When responsibility boundaries are unclear, the release pipeline devolves into a string of pretty status labels: Generated, Checked, Synced, Completed. When things actually go wrong, no one can pinpoint which step failed to provide the necessary evidence.

SFD’s recent daily update anomalies offer a concrete example. Multiple streams—diaries, articles, science popularization, and skills—were running simultaneously. However, a task returning “ok” didn’t mean the content actually appeared on the live page; a cover image returning a 200 status didn’t guarantee the detail page truly referenced it; and daily content publication didn’t ensure the theme changed every day.

Therefore, small teams don’t need more role titles; they need stricter division of responsibilities.

The Publisher is solely responsible for delivering content to the CMS and recording the ID, slug, locale, and publication time. It should not casually declare that quality checks have passed.

The Reviewer is solely responsible for evaluating the topic, body text, and duplication risks. It must be able to articulate how the current piece differs from the previous seven. If it can only say “the expression is more complete,” it does not pass.

Visual QA is solely responsible for what is actually visible on the page: whether the list page displays the cover, whether the detail page references the cover, and whether mobile cropping is excessive. It should not merely check for file existence.

The Operator is responsible for the final escalation decisions. Humans must determine which issues can be fixed on the spot, which require pausing automated publishing, and which should involve overwriting modifications while preserving original links.

This division of labor may seem cumbersome, but it addresses the most costly problem for small teams: never let a single AI Agent act simultaneously as author, reviewer, publisher, and acceptor. The fewer the roles, the harder the evidence must be; the more automation there is, the less responsibilities should be mixed.