Why our AI agents can't publish — the design decision behind draft → approve

The most common question we get is some version of 'why can't I just let the agent post?' Here's the reasoning, because it shaped everything else in Nimply.

Capability isn't the bottleneck any more. A model can draft a reply in your voice, judge whether a comment should be hidden, or lay out a carousel well enough to put near a customer. Adoption still lags, and the reason is trust: nobody hands software their public voice without proof it can't act outside the limits they set.

So we made the guardrails the product:

• Every agent run produces a draft. Drafts go through your approval workflow before they reach the scheduling pipeline — the same rail a human-written post uses.

• Auto-reply is off by default. An agent proposes; you (or a teammate with approval rights) send.

• Any agent can be paused per thread, and a human taking over a conversation makes the agent stand down.

• Disconnecting a channel automatically pauses the agents attached to it.

• Every run is metered in credits you can see, priced against real provider cost, so there's no surprise bill and no mystery about what a run cost.

What this costs us: real publishing infrastructure, an approval workflow, and vendor-approved app review with every platform we publish to. That's months of work a thin wrapper can't retrofit, which is also why we think it holds up as a moat.

Where it's honestly hard: volume. If you're clearing hundreds of replies a day, approval becomes the job. Our answer so far is batching approvals, letting trusted teammates approve, and making the agent's confidence visible so you can skim. If you've solved this differently, I'd like to hear it.