Most businesses that let an AI agent touch money do the sensible thing: they put a person in the loop. The agent drafts a refund, a purchase order, or a credit, and a manager approves it before anything happens. That approval step is the main reason owners feel safe letting an agent near the checkbook.
But an approval is only as good as its timing. Picture this: a customer asks for a $500 refund. The agent prepares it and sends the request to a manager. The manager is out, so a colleague handles the customer by phone and refunds them directly. Two days later the manager catches up on approvals and clicks yes. Does the agent now send a second $500?
Full disclosure: I built Agent Ready Rankings (https://rankings.leffke.com), a site that scores AI agent tools out of 10. The scores below are overall scores from my November 11 data run, not a same-day refresh. Browsing the site is free; the full score breakdown is part of Pro, $10/month.
The stale "yes" problem
Builders call these approval steps interrupts: the agent pauses, waits for a human answer, and then resumes. The risk is that the answer arrives after the world has moved on, or that one answer gets matched to the wrong paused request. Either way, the agent does exactly what it was told, and you still lose money.
This isn't exotic. Anything with a queue of approvals and busy people will eventually see late answers. The question is whether your agent notices.
The framework underneath matters
Your builder picks a framework that runs the agent's steps, including those pauses. Some are adding guardrails here. LangGraph's 1.2.20 release, published November 4, 2026, added a separate id for each interrupt and a strict resume mode that refuses stale interrupt answers. That doesn't make any agent safe on its own, but it gives your builder a tool to stop a late yes from triggering an old action.
The agent frameworks I track (overall scores)
LangGraph: 8.0 out of 10. Stateful orchestration from the LangChain team, built around graphs, checkpoints, and durable runs. Free to use, with paid options. Latest tracked release 1.2.20 on November 4, 2026.
CrewAI: 6.85 out of 10. A framework for setting up a "crew" of agents with different roles. Free to use, with paid options. Actively updated; latest tracked release 1.15.35 on November 3, 2026.
Microsoft AutoGen: 5.65 out of 10. Free and capable for multi-agent conversations, but the last release I've tracked was September 30, 2025. If your builder uses it, ask how they keep it current.
You won't pick the framework yourself. But you can ask the questions that tell you whether your builder has thought about this.
A short owner checklist
Before an agent is allowed to spend, refund, or order anything, ask:
- Which actions need a human approval, and is there a dollar limit above which they always do?
- Does an approval expire, and what happens if it comes back late?
- Is each approval tied to one specific request, so a yes can't be applied to the wrong one?
- If the situation changed while the agent was waiting, does it check again before acting?
- Is every approval recorded, with who said yes and when?
If the answers are vague, the approval step is giving you comfort, not control.
About the rankings
Agent Ready Rankings covers 22 AI agent tools across coding agents, LLM APIs, agent frameworks, voice, memory, evaluation and safety, and deployment and observability. I built it, so treat every score as one informed opinion and tell me where you think I'm off.
Agent frameworks category: https://rankings.leffke.com/categories/agent-frameworks
