An agent that can act is more useful and more dangerous than one that can only talk. The talking kind drafts an email. The acting kind sends it, refunds the order, and deletes the row. The gap between those two is where most teams get nervous, and they should.
The fix is not "keep a human in the loop" as a slogan. The fix is deciding which specific actions stop and wait for a person. That decision is design work. Done well, the agent ships real output. Done badly, you either approve everything by reflex or block nothing that matters.
Read is free, write is not
Start by splitting actions into reads and writes. Reads are cheap and reversible. An agent querying your database, pulling a ticket, or summarizing a thread cannot hurt you. Let those run unattended.
Writes change the world. Sending a message, charging a card, updating a record, calling an external API with side effects. These deserve scrutiny. Not all of them, but this is where you look.
Most teams over-gate reads and under-gate writes. They make a person watch the agent think, then rubber-stamp the one action that actually matters. Flip it. Let the reasoning run. Gate the consequence.
Gate by blast radius, not by step
The wrong model is "approve every step." That trains people to click yes without reading. After fifty harmless approvals, the fifty-first gets the same reflex click, and that one wired money.
Gate on blast radius instead. Ask two questions of each write. Is it reversible? How many people or dollars does it touch?
A refund under $50 to one customer is low radius and reversible enough. A bulk update across every account is high radius and hard to undo. Tier the gates. Auto-run the small stuff. Stop and ask on the big stuff. Require two approvals on the irreversible stuff.
This keeps the gate meaningful. People only get interrupted when interruption is earned, so they actually read the prompt.
What a good gate shows
A gate is only as good as what it puts in front of the human. "The agent wants to proceed" is useless. Approve what, exactly?
Show the literal action. The exact API call, the exact recipient, the exact dollar amount, the exact rows. Show what the agent believes will happen and why it chose this. If a person cannot reconstruct the decision in ten seconds, the gate failed before anyone clicked.
We build this into the agents we ship. A support agent we built answers a customer in 12 minutes, down from 4 hours. It drafts and acts fast, but the actions with teeth surface a clear, specific approval before anything fires.
Log the approvals too
Every gated action should leave a record. Who approved it, when, what the agent proposed, what actually ran. This is not bureaucracy. It is how you find the gate that was too loose and the gate that was pure friction.
After a month of logs you learn which approvals were always yes. Those are candidates to automate. You also learn which agent proposals were quietly wrong, which tells you where the model needs better guardrails. The log turns gate design into something you tune with evidence instead of fear.
The goal is fewer gates over time
A good approval system shrinks. You start cautious, gate broadly, watch the logs, and remove gates that never caught anything. The agent earns autonomy by behaving.
That only works if you can measure it. If approvals are a Slack message someone eyeballs, you have no data and the gates live forever. Make them structured, logged, and tiered from day one. Then let the boring ones retire.
This is the same discipline we bring to everything we ship. We are a 2-person team, ex-AWS Federal SA background, that put 10 production products into the world in 7 months. Agents that act in production are most of that work now, and approval gates are why clients trust them with real systems.
If you are putting an agent near anything that sends, charges, or deletes, gate design is the part to get right first. OpenClaw Installation is our fixed-price way to stand up an agent with these gates built in from the start. Fixed price, you own the code, we exit. See pricing for where it lands.
Want a second set of eyes before you wire an agent into production? Book a free 30-minute call and we will walk through where your gates should go.