The pitch is right about the destination and dangerously casual about the route. An agent that acts on what it heard in a meeting is acting on the messiest input in software: half-finished sentences, sarcasm, thinking out loud, and speech-to-text that regularly hears "Acme" as "acne." The question that will define this category is not whether meeting agents should act. They should. It is whether they act on their own judgment or on your yes.
We build Harlyn, and we chose the yes. Every action that touches your systems is staged, read back, and executed only on explicit approval. Here is the case for that being the right architecture, not just the cautious one.
Read any real transcript and you will find the problem immediately. "We should probably push the close date" is followed two minutes later by "actually, let's hold off." A speaker says "I'll take care of the renewal" meaning next quarter, or meaning hypothetically, or quoting what the client said. Someone jokes "just delete that account." A speech-to-text hiccup turns a customer's name into a different customer's name.
A human assistant sitting in that meeting handles this effortlessly, because a human assistant confirms: "just to check, you want the close date moved to March?" That confirming instinct is what most automation quietly deletes, because confirmation is friction and friction demos badly.
So the extraction-only products (the note-takers) hand you a list of action items and do nothing, which is safe and unsatisfying. And the automation-first products act on their interpretation, which is satisfying right up until the interpretation is wrong inside your CRM, where wrong answers are quietly durable: a bad close date flows into the forecast, and the forecast flows into decisions made by people who never heard the original sentence.
The approval gate is a simple contract: the agent may hear, interpret, and prepare anything, but it may execute nothing until a human says yes to a specific, stated plan.
In a live meeting it sounds like this. Someone says "Harlyn, take care of those." The agent reads back what it is about to do: which record, which field, which change. Someone in the room says yes. It executes, confirms out loud, and the whole exchange takes seconds. If the read-back is wrong, the room corrects it before anything happened, which is the entire point: the mistake got caught at the cost of one sentence instead of a forensic cleanup.
When nobody in the room can approve, because the agent is attending as a delegate for someone absent, the staged actions wait in that person's dashboard. They review the list after the meeting, approve with a click, and only then does anything run. The click is the approval; there is no timer that runs out and acts anyway.
Two details matter more than they look. First, the read-back names the actual change, not a summary of a summary; "update Acme Renewal close date to March 15" can be verified by ear, "handle the follow-ups" cannot. Second, an approval covers exactly what was read back. Anything discovered mid-execution goes back through the gate rather than riding along on the earlier yes.
The honest objection: if a human approves everything, how is this automation?
Because the expensive part of follow-through was never the click. It is remembering, assembling, and doing. The agent caught the action item while the conversation moved on, resolved it against the right record in your CRM, drafted the exact change, and held it ready. What is left for the human is the one thing humans are actually good at in this loop: recognizing whether the stated intent matches what was meant. Five seconds of judgment on top of zero minutes of work.
There is also a compounding reason: the gate is what lets you delegate more, not less. Teams give real system access to an agent they can predict. An agent that might do anything with your CRM gets read-only access forever, sandboxed by fear. An agent that provably does only what was approved gets invited deeper, because every yes it has ever executed is a receipt. Trust in agents is not built by capability. It is built by predictability, and the gate is predictability made mechanical.
The transcript becomes an audit trail too: what was staged, what was read back, who said yes, what ran. When a number in the CRM looks odd three weeks later, the answer to "why did this change?" is a recorded sentence with a recorded approval, not a shrug at the automation.
We would rather this were boringly universal than a differentiator. If you are evaluating AI meeting agents this year, put these five questions in the RFP:
Harlyn's answers are no, yes, yes, nothing, and yes. In private beta, free, and the approval gate is not a settings toggle. It is the architecture, because an agent acting in your meetings is acting in your name, and your name should be spent by you.