Agents

An Agent makes case-by-case judgment calls — scoped, always gated by an approval wherever the decision acts outward. Agents live inside Workflows, never floating freely alongside them.

Reply agent

The reply agent reads incoming replies and decides: interest, decline, follow-up question, meeting request? It proposes the next action (reply, suggest a time, stop the sequence) — every outward-facing action (a sent reply, a booked meeting) runs through an approval wherever it isn’t yet learned as safe. Every classification flows back into your Knowledge as evidence for the next learning loop.

Closed-loop agent

A quartet of four roles that continuously observes and re-tunes your campaigns:

• Observer (every 15 min) — classifies new replies, extracts performance signals.
• Optimizer (every 6 hours) — derives decision proposals from them: switch copy, adjust cadence, sharpen a playbook.
• Executor — only implements approved decisions.
• Measurer (daily) — measures the actual outcome after 48 hours and adjusts the confidence of the underlying learnings.

Proactive agent

Reaches out instead of waiting to be asked: morning briefing, follow-up alerts, nurture reminders, meeting prep — through Command or as a reminder, never as an automatic send without your approval wherever an action acts outward.

Shared principle: guardrails instead of autonomy

No agent in this system pauses or escalates a campaign autonomously — it proposes, in the feed or in chat, and you decide. Every agent’s scope is explicitly bounded: a reply agent evaluates replies, it doesn’t change playbook structure; that’s handled by an approval for structure-changing proposals (see Approvals).