Use a different AI model to double-check your work
I use a different model for a second opinion on my AI agent's work, and it's been generally good so far.
I got the idea from Amp Code's Oracle, where you can call in a stronger model to double check what you're doing. I call mine advisors.
When I want a second opinion, I call in the strong version of the other model. So currently, I'm using Anthropic Opus as my daily driver, and the advisor would be OpenAI Astra. Also, I've had better results with OpenAI for web search, so for things that require web search, I'll call in the advisor.
I think it works mostly because it's a fresh context, but also it being a different model helps a bit, I think. There's a bit of diversity-of-thought going on as well.
Also, I've noticed that models generally prefer their own family, so by default, I use the other model family. Sometimes, I'll call in both advisors (other model family + own model family). If they disagree, that's interesting, and I dig in. I also try to get them to converge, but maybe converge isn't good, they should spar.