Autonomous social media agents

The challenge

A group of cafe and restaurant chains in Southeast Asia needed a steady stream of social content for every location they operated, across multiple regions and languages. The traditional answer is a content team per market: copywriters who know the local language, the local promotions, and the local audience. That model does not scale for a multi-brand food and beverage operator. In practice, head office marketers were spending their weeks hand-writing posts, translating them, and pushing them out location by location, while the actual cadence and quality of posting varied wildly between markets.

The brief was direct: keep every location’s feeds active with content that sounds like the brand and reflects what that specific cafe is actually selling this week, without hiring a content team for each market and without letting an unsupervised model post on the brand’s behalf.

What we built

We built an autonomous content pipeline that runs from brand inputs to published posts, with exactly one human checkpoint in the middle.

Draft and localization agents

The pipeline starts from three structured inputs: the brand guidelines, the current menus and promotions, and a content calendar. A draft agent consumes these and writes candidate posts that are tone-aware and offer-aware, so a post about a seasonal drink references the real promotion running at that location, in the brand’s actual voice. A localization agent then adapts each draft per region and language, rather than running a flat translation. The same campaign lands differently in different markets, and the agent adjusts phrasing, references, and language accordingly.

Brand safety review gate

Every localized draft passes through a human-in-the-loop review gate before anything is published. This gate exists because the failure mode for autonomous publishing is asymmetric: a missed post costs little, while an off-brand or tone-deaf post on a public channel costs real trust. Reviewers approve drafts or send them back with notes, and rejected drafts loop back to the agents for revision. The agents have full autonomy to draft, localize, and revise. They have no autonomy to publish. Nothing reaches a public channel without passing the gate. That division turned the marketing team’s job from writing everything to reviewing a queue, which is where the 15 hours per week of freed-up manual work comes from.

Scheduled publishing across platforms and locations

Approved posts flow into a publishing scheduler that maintains a queue per channel. Content goes out to Instagram, Facebook, and TikTok on a per-cafe-location basis, so each outlet’s feeds stay active on their own cadence without anyone manually posting.

Engagement feedback loop

Published posts report likes, reach, and click-through back into an engagement metrics store, and those metrics feed back to the content agents. Formats and angles that perform in a given market inform what the draft agent produces next for that market. This closed loop is what moved the numbers over time, ultimately reaching 3.5x follower engagement.

How it was delivered

We delivered the system incrementally rather than as a single launch. The draft and localization agents came first, running behind the review gate from day one so the client’s team could judge output quality on real campaigns before any automation touched a live channel. Once approval rates were consistently high, we wired up scheduled publishing per platform and location, then closed the loop with engagement feedback. Throughout, the client’s marketers stayed in the workflow as reviewers, which kept brand judgment where it belongs and gave us a continuous source of correction signal for the agents.

What shipped

• Draft agent producing tone-aware, offer-aware posts from brand guidelines, menus and promos, and the content calendar
• Localization agent adapting every post per region and language
• Human-in-the-loop brand safety review gate with a revise-and-resubmit loop, the sole path to publication
• Scheduled publishing with a queue per channel, covering Instagram, Facebook, and TikTok for each cafe location
• Engagement feedback loop routing likes, reach, and CTR back into the agents’ drafting decisions

The result is localized content on autopilot: 3.5x follower engagement and 15 hours per week of manual work returned to the marketing team, with a human still signing off on every post.