machines, runtimes, and the layer under agents
an agent is only as real as the machine it runs on. everything interesting i've hit in the last year was a hardware fact wearing a software costume — codesigning, a simulator that only exists on real silicon, a scheduler that can't be told the truth about capacity.
sdk / cli / agent │ ▼ ┌────────────┐ ┌──────────────────────────┐ │ control │───────►│ your mac │ │ plane │◄───────│ xcode · simulators │ └────────────┘ │ codesign · homebrew │ │ │ real safari, real keychain │ │ └──────────────────────────┘ ▼ fleet: n machines, parallel, stateful
the part that can't be faked in a vm is the part worth building on. codesigning and the simulator are the whole reason this has to be a real mac, and they're also why a fleet is stateful — machines are not interchangeable once they hold identity.
real macOS as a cloud primitive. the constraint isn't compute, it's identity: signing, keychains, and what a machine is allowed to be. more →
one mcp connection as the memory layer, so context and procedures are a property of the workspace rather than of a session. more →
whether fleet scheduling for agents converges on batch-scheduler shapes from twenty years ago, or whether long-horizon breaks the analogy.