What you'll be able to do after this guide: put a standard set of checks in front of every outreach action, choose between human-sent and automatic sending, and respect both the person whose account is used and the platform's rules.
Gates, cheapest first
Every outreach agent runs the same checks before acting, in order of cost. A candidate stopped by an early check costs nothing: no browsing, no drafting.
1. Status gate. The company is at Review (or Contacted, for a follow-up) and has a research date. This is a CRM read only.
2. Deal gate. No open deal past its first stage. If a person is already in a conversation, the agent stays out. Check deals linked to the company and to the person, because duplicate company records mean a deal can hang off the "other" one. If the deal lookup fails, skip the candidate. A failed check must never be read as "no deal".
3. Reply check. If the prospect has ever replied, stop and flag them for a person. Not "replied since our last message": any reply at all. Don't snooze them; a snooze would hide them from the person who needs to answer. The agent checks replies the CRM knows about: synced emails and replies a person has logged.
4. Recent conversation. If there's any message in the last 30 days, from anyone, wait. This catches messages a person sent by hand. Where the agent can't see the conversation (on a social platform, for example), this check and the reply check become the first line of the person's task: "Before sending, check the thread. If they have replied, or anyone has messaged in the last 30 days, don't send; note why and complete the task."
5. Cooldown. At least 14 days since our last message, keyed on the person's profile URL rather than the CRM record, so duplicate person records can't slip a second message through.
6. Claim check. Re-verify any fact the message will mention, and confirm every link opens. Facts go stale: a tool confirmed in research can be gone a few weeks later. A page you planned to publish may never have shipped.
Only then does the agent act, and it records the action straight away, before moving to the next candidate. Deferring writes to the end of a run leaves a window where a message has been sent but the cooldown doesn't know about it.
When in doubt, the answer is "don't send"
Put this sentence in every outreach spec: "Never resolve a conflict in the direction of sending a message." If the spec and the prompt disagree, the more restrictive one wins. If a check can't be completed, the candidate is skipped.