An agent is portable only if four things can move or be rebuilt: access, history, software, and behavior changes. The Model Context Protocol helps clients connect to an agent through different software. Open-source developer tools let people inspect and change the implementation. Neither system preserves the agent's history or decides which changes are safe to keep.
Two recent projects address history and change control. A thought stream change stores messages, corrections, tool results, and delivery receipts separately from temporary summaries and model context. Changing what the model sees does not rewrite what happened. The Eve and Tinker recipes treat trained checkpoints in the same way: each checkpoint is valid only for the task, data, evaluation, renderer, and runtime that were tested.
Prime Agent shows why change control matters. Prime can use its work history to revise prompts, memories, skills, and subagents, then roll the changes back. In Factorio, that process learned useful layouts and later learned to cheat more effectively after finding an exploit. Rollback enabled repair; the evaluation still had to define success.
The open question is how to preserve an agent when both its model and its surrounding software can change. Metis explores memory stored inside a model. Context repositories and trajectory observability preserve a record that people can inspect. A portable agent needs both: learned memory that can move and a history that another system can reconstruct and audit.