I don't know for this project specifically, but previous contributions to this project perhaps have turned "author rights" over to the project, then the project relicensing is within their rights. Projects often do this because these license issues have come up before.
You can do this when you publish software under the GPL, and then the FSF will take care of enforcing the license for you. It's a popular thing to do because if you don't, you'd have to get your own lawyer, etc.
the parallel comment about MIT licenses allowing relicensing is also valid, just rounding out the explanations.
What you're talking about is called a CLA, a contributor licensing agreement, and is precisely what the project is boasting they are not doing. This prevents the company from ever changing the license from AGPL to something else from now on, since they wouldn't own enough of the copyright.
Not quite. The previous commenter was referring to how the FSF handles contributions—copyright assignment—not a CLA. Your conventional CLA doesn't transfer copyright ownership to the upstream project. It just permits them to do a lot more with the contributor's work (including e.g. arbitrarily relicensing) than other recipients/participants get to do, without actually making the original author give up ownership.
Hypothetically, if 95% of code is from the company and 5% is from community, would it be possible for company to remove commits made by community, reimplement affected parts and move the project from AGPL to closed source?
I'm talking about what is published on the GNU and FSF sites. Whether they still actively do it is an interesting question and I welcome your feedback, but I'm not spreading false ideas, and my comment is "education" about various aspects of licensing.
You can do this when you publish software under the GPL, and then the FSF will take care of enforcing the license for you. It's a popular thing to do because if you don't, you'd have to get your own lawyer, etc.
the parallel comment about MIT licenses allowing relicensing is also valid, just rounding out the explanations.