Updated · By the Nexus team
Put a human approval gate in front of actions that are irreversible, customer-facing, or expensive: sending external messages, publishing, merging or deploying code, deleting data, spending money and changing permissions. Let everything else run on its own, and log all of it. A gate should hold the finished output for review, not interrupt the agent on every step.
Gate the action, not the agent
Asking for approval on every step makes an agent slower than doing the job yourself. Asking for none makes it unsafe to give it real work. The practical line is the action: research, drafting and analysis run freely; anything that leaves the building or cannot be undone waits.
A starting policy
| Action | Default | Why |
|---|---|---|
| Research, reading files, drafting | Runs freely | Nothing leaves your control |
| Sending email or messages outside the company | Needs approval | Speaks in your name |
| Publishing to a website or social account | Needs approval | Public and hard to retract |
| Merging or deploying code | Needs approval | Affects production |
| Deleting data | Needs approval | Irreversible |
| Spending above a set amount | Needs approval | Money |
| Changing access or permissions | Needs approval | Security |
Loosen a gate once an agent has a track record on that kind of job. Tighten it the first time something slips through.
What a good approval looks like
- The finished output, not a question. Review the email, not "may I write an email?"
- The context: which job, which agent, what it cost so far, what it read.
- Approve, edit then approve, or reject with a reason the agent can learn from.
- A record: who approved what and when, kept in the audit trail.
- A way to step in. Sometimes the fastest fix is taking over the run for a minute.
How Nexus does it
Nexus holds output that crosses an approval gate until a person signs, records every change in an audit trail, and lets a human join a live agent session to take the wheel mid-task. Gates sit alongside per-agent budgets checked before the spend.
Frequently asked questions
What is human-in-the-loop for AI agents?
A design where an agent works autonomously but pauses at defined points for a person to review, edit or reject its output before the action takes effect.
Do approval gates slow agents down?
Only at the gates. If gates sit on irreversible or external actions, most of a job runs without waiting and the review happens once, on the finished result.
Why do AI agents need an audit trail?
To answer what happened, who allowed it and what it cost, for debugging, for customers and for compliance. Without one, a mistake by an agent is very hard to trace.
Keep reading
Try Nexus free
Hand an agent one real job under a budget you set. The free plan never expires. Start on the web or see pricing.