Pick on the constraint that will not move. This decision outlasts whichever coding agent you are using this month.
You want the agent to change the content model, not just fill it in. Sanity. MCP reaches schema and releases as well as documents, the Studio schema is code, and TypeGen tells your front end what a query returns rather than only what a document contains.
You want the whole system in one repository with no vendor database. Payload. The model, the admin panel and the API are your application. Accept that you will run the installer yourself, once.
Your team is small and shipping marketing pages. Prismic. The vendor has pointed its own documentation at CLI-plus-agent as the way to work, the model pulls down as files with types, and the pricing ladder has four published steps rather than a cliff.
You want MIT licensed and self-hosted, with the model in files. Strapi. The schema is JSON an agent can write, the plugin ecosystem is the largest here, and cloud pricing is per project rather than per seat.
You are on Laravel and want content in git alongside code. Statamic. Flat files are the purest form of what this list rewards, on the understanding that you are wiring the agent up yourself.
You have a production database you are not moving. Directus. It sits on what you already have, and the snapshot file makes the model reviewable even though the database holds it.
Non-developers own the content model and always will. Contentful. The GUI needs no toolchain and no repository access, which is exactly why large organisations keep choosing it, and the migration CLI is there when engineering needs the model in code.
Your editors need to see the page while they work. Storyblok. Rank it eighth here and second on any list about editing, because both are true and they are different questions.
Several systems need to sit behind one API. Hygraph. Content federation does something nothing else here does natively, and Early Access on MCP is a timing problem rather than a permanent one.
If none of those is binding, the useful test is smaller than the shortlist suggests: open a terminal, and see how far an agent gets before it needs you. That single check separates these nine faster than any feature table, and it costs an afternoon rather than a migration.