What is approval fatigue?
Approval fatigue is the tendency of people asked to approve an AI agent's actions too often to stop reading the requests and approve by habit, which makes frequent permission prompts less safe than they appear.
Asking a person before every agent action sounds like the safe default. In practice, the tenth identical prompt in an hour gets approved without a glance, and the one prompt that mattered goes through with the rest.
Anthropic's October 2025 sandboxing post names the problem: constant approval requests can lead to approval fatigue, where users stop paying close attention to what they approve. A gate that is always passed is not a control.
What the evidence shows
Anthropic's February 2026 study of agent autonomy found that experienced users auto-approve more but also interrupt more: oversight shifts from approving each step to monitoring and stepping in. The same study describes a deployment overhang, in which models can handle more autonomy than people currently give them, and recommends investment in post-deployment monitoring.
Sandboxing cut permission prompts by 84% in Anthropic's internal use, by making most actions safe to run without asking. Claude Code's auto mode goes further, with classifiers approving or denying actions, and Anthropic publishes its limits: a 0.4% false positive rate and a 17% miss rate on overeager actions, with the note that it does not replace careful human review.
Designing approvals people actually read
Approve by risk tier
Low-risk, reversible actions run inside a sandbox without a prompt; only high-impact actions ask. OpenAI's agent guide ties human intervention to failure thresholds and high-risk actions.
Show consequences, not commands
A preview of the email, the refund amount or the diff is easier to judge than a raw tool call.
Batch related steps
One approval for a reviewed plan beats ten for its individual steps.
Pause without cost
Google's long-running agent guidance describes delegated approval, where an agent waits for a person without using compute, so approvals can wait for the right reviewer.
Measure the gate
If nearly every request is approved within seconds, the prompt is not being read and should be removed or redesigned.
How Motiontic's engineers apply it
Motiontic's engineers set the containment boundary first, then add approvals only for the few actions that remain irreversible or costly inside it. Approval rates and response times are logged, and a gate that is approved almost automatically is treated as a design problem rather than as evidence of safety.
Three oversight models compared
| Approve every step | Risk-tiered approval | Sandbox plus monitoring | |
|---|---|---|---|
| Prompts per task | Many | Few | Rare |
| Approval fatigue risk | High | Low | Very low |
| Depends on | Reviewer attention | Accurate risk ratings | A well-defined boundary and good traces |
| Fits | Early pilots with few runs | Most production agents | Agents with contained, reversible actions |
Common questions
What is approval fatigue in AI?
Approval fatigue is when people asked to confirm an AI agent's actions too often start approving without reading, so the approval step stops catching the actions it was meant to stop.
How is approval fatigue avoided?
By asking less often and about the right things: sandbox routine actions so they need no approval, reserve prompts for irreversible or high-impact actions, and show a clear preview of the consequence.
Does human in the loop cause approval fatigue?
It can, when every step needs a person. Human in the loop works when approvals are rare and meaningful, which is why many teams move routine steps to monitor-and-intervene oversight.
Related
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
Human in the Loop vs Human on the Loop
Approving each step, compared with monitoring the run and stepping in.
Agent sandbox
Hard walls around where an agent can read, write and connect.
Blast radius (AI agents)
The worst an agent could do, measured by what it can reach.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Classifier-gated permissions (auto mode)
A model reviewer in place of a permission prompt, with published error rates.
Sources
- Anthropic, Beyond permission prompts: Claude Code sandboxing (2025-10-20)
- Anthropic, Measuring AI agent autonomy in practice (2026-02-18)
- Anthropic, Building Claude Code auto mode (2026-03-25)
- OpenAI, A practical guide to building agents (PDF)
- Google Cloud, Five must-have guides to move agents into production (2026-05-06)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.