LotsAgentWhich actions should your AI agent run autonomously, and which need a human to approve them first? The answer depends on the risk, the context, and how much trust you have built over time — but there are clear decision rules that make this practical rather than philosophical.
Anthropic's February 2026 research on agent autonomy found that most agent actions in production are low-risk and reversible — software engineering alone accounted for nearly 50% of agentic activity on their platform. But the same research showed that the remaining actions touch domains — healthcare, finance, cybersecurity — where a wrong call can have real consequences. That is exactly where review gates matter.
This article gives you a practical framework for deciding which actions run autonomously, which require human approval, and how to configure review steps that scale as your operation grows.
The Review Step Spectrum
Not all agent decisions need the same level of oversight. Think of review gates as a spectrum with four zones:
- Fully autonomous. The agent acts without any human check. Suitable for low-risk, high-frequency, reversible actions: reading messages, categorizing tickets, drafting responses, running software builds.
- Draft and approve. The agent prepares an output and presents it for human review before executing. Suitable for customer-facing communication, financial recommendations, and actions that represent the organization externally.
- Flag and escalate. The agent detects an exception — an unusual request, a permission boundary, a high-cost action — and routes it to a human with context. The agent continues working on everything else.
- Human-only. Specific actions are locked behind a human-only gate. No agent can execute them regardless of permissions. Suitable for irreversible operations: deleting accounts, modifying access controls, approving large payments.
Most agent platforms, including LotsAgent, support all four modes natively. The configuration question is which actions go where.
Decision Rules for Review Gates
Four criteria determine where an action falls on the review spectrum. They form a consistent decision framework that you can apply to any agent workflow:
1. Money
Does the action move, commit, or authorize funds? Sending an invoice, issuing a refund, updating a subscription tier, or approving a discount all touch money. The review requirement scales with the amount: small, frequent transactions might run autonomously with a daily audit; anything above a threshold needs approval.
2. Identity
Does the action create, modify, or delete a user account or change permissions? Identity changes have a long tail of consequences. A single misapplied permission can grant access to systems the user should never see. These actions should default to human approval with very narrow exceptions.
3. External communications
Is the agent speaking to a customer, partner, or regulator? Draft-and-approve is the standard here for a simple reason: the organization owns the message. Autonomous replies can work for internal triage or status updates within a controlled scope, but anything customer-facing should pass through review until the agent has a proven track record.
4. Irreversibility
Can this action be undone? If not, it needs a human gate. Deleting data, terminating infrastructure, changing production configurations — these actions should never run autonomously. The more irreversible the action, the higher the review requirement.
Setting Thresholds That Make Sense
Decision rules without thresholds are just intentions. To operationalize them, set concrete boundaries:
- Dollar thresholds. Refunds under $50 run autonomously; $50–$500 need single approver; over $500 needs two-person approval.
- Recipient types. Internal messages auto-send; customer-facing messages draft-and-approve; regulatory communications human-only.
- Content categories. Scheduling posts auto-approve; product announcements draft-and-approve; pricing changes human-only.
The exact numbers depend on your business, but the principle is universal: thresholds convert policy into enforceable rules.
How Review Steps Work in LotsAgent
In LotsAgent, review steps are configurable per agent, per tool, and per action type. You do not write custom code or add middleware. Every agent has identity, execution logging, and configurable review gates built in.
Here is how a typical review configuration looks for a support agent:
- Reading and categorizing incoming tickets. Fully autonomous. The agent reads the message, categorizes it by topic and urgency, and assigns it to the correct queue.
- Drafting a response. Draft-and-approve. The agent prepares a response based on the knowledge base and presents it for human review. The human edits or approves before it sends.
- Processing a refund. Flag and escalate. The agent detects the refund request in the message, checks the amount against policy, and surfaces it in an approval queue with context. The human authorizes or rejects through the dashboard.
- Changing a user's account tier. Human-only. The agent never has permissions to modify subscription or access levels. The request is forwarded to an administrator workflow.
All of this is configured through LotsAgent's interface — no code, no YAML, no infrastructure setup. The audit trail logs every decision, every tool call, and every human override automatically.
The Trust-Building Cycle
Review gates are not static. As your agent demonstrates consistent, safe behaviour, you can expand its autonomy. The pattern is cyclical:
- Start strict. New agents run with tighter review gates. Every meaningful action gets a human check. For a detailed walkthrough of this progressive approach, read The 7-Day Production Rollout: From First Agent to Unattended Run.
- Gather evidence. The audit trail shows what the agent does, how often it needs corrections, and where it makes good independent decisions.
- Expand boundaries. Based on the evidence, move specific actions from draft-and-approve to autonomous, or raise dollar thresholds.
- Monitor drift. Agents drift over time — model updates, tool changes, or shifting workflows can change behaviour. Periodic audits catch this before it becomes a problem. Build AI Agents That Don't Break: 8 Production Failure Modes covers exactly how to monitor and catch this kind of drift.
This is consistent with what Anthropic's research found: experienced users auto-approve more frequently as they build trust. The platform needs to support that transition with audit trails and configurable gates.
Common Review Step Configurations
Here are three common configurations that map to real agent types. If you are building a support workflow specifically, How to Turn a Support Inbox Into an Agent Workflow That Actually Escalates the Right Things goes deeper on the full setup.
Support Agent
- Ticket triage: autonomous
- Drafting replies: draft-and-approve
- Refund requests under threshold: autonomous with daily audit
- Account changes: human-only
Sales Agent
- Lead qualification: autonomous
- Follow-up emails: draft-and-approve
- Proposals and quotes: flag and escalate with dollar thresholds
- Contract changes: human-only
Operations Agent
- Scheduled data syncs: autonomous
- Alert triage: autonomous with escalation rules
- Infrastructure changes: flag and escalate
- Production deployments: human-only
The review step is what separates accountable agent operations from black-box automation. Set the gates where the risk is, start strict, and expand only when the evidence supports it.
Create your first agent free at lotsagent.com.
FAQ
What determines whether an agent action needs human approval?
Four criteria: whether the action involves money, changes identity or permissions, communicates externally, or is irreversible. Actions that trigger any of these criteria should default to at least draft-and-approve review until the agent demonstrates consistent safe behaviour.
Can review gates be adjusted after the agent is deployed?
Yes. Review gates are configuration settings, not code changes. You can move actions between autonomous, draft-and-approve, flag-and-escalate, and human-only as the agent builds a track record. The trust-building cycle — start strict, gather evidence, expand boundaries, monitor drift — is the recommended approach.
What happens when an agent action is flagged for review and nobody approves it?
The action does not execute. Most review gate implementations, including LotsAgent, log the pending action, notify the assigned reviewer, and escalate to a secondary contact if not acted on within a configurable time limit. The agent continues working on other tasks unaffected.
Do review gates slow down workflows significantly?
For high-frequency, low-risk actions running autonomously, there is no delay. For actions requiring review, the delay depends on how quickly reviewers respond. Draft-and-approve mode is the most common compromise: the agent does the hard work of preparing the output, and the human makes a quick decision.