Runtime vs model switching

It is easy to assume Constant swaps a model. It does not. It swaps the whole runtime underneath one conversation. The difference is the whole point.

Model switching

Picking GPT vs Claude inside one app is model switching: same harness, same tools, same session store, you just change which weights answer. The application and its scaffolding stay put.

Runtime switching

Constant swaps the entire agent CLI. Codex, with its tools, prompts, session format, and resume behavior, becomes Claude Code, with its own. The one thing held constant is the conversation transcript. Everything else, the harness and the model and the character, changes together.

Memory vs mind

There are two layers:

• The transcript is memory: what was said. It is portable, and it is what Constant carries.
• The weights are the mind: the character baked in by training. Not portable, and exactly why Claude answers differently than Codex on the same thread.

Carry the memory; let the mind change. Same notebook, different brain, different answer. The divergence is the value, not a bug.

Why it is even possible

Because agents are stateless between turns and reconstruct the conversation from a file each turn. There is no live mind to transplant; there is a file, re-read each turn, and Constant writes that file. See How it works.

What it means in practice

• The two runtimes will disagree, take different approaches, and have different strengths. Use that on purpose.
• The conversation continuity is seamless; the agentic substrate (tools, reasoning) does not cross. See What carries.