Frameworks
The framework is the harness the agent's code lives in: the thing that owns its loop, its tool wiring and its state between steps. Calafia writes the agent for you and this axis decides which shape that code comes out in.
What it changes is mostly ownership, not intelligence. The same goal, model and tools produce broadly the same work in any of them. What differs is where the agent can run, what it is idiomatic to extend, and who is holding it a month later — an exported project is yours to change, and stops tracking anything we do afterwards.
It does have real behavioral edges. Frameworks differ in how they handle a failed tool call, how much of the loop is theirs versus the model's, and how a long-running step is checkpointed. Those show up under load rather than in a first run.
Frameworks in this deployment
What this deployment can export or host today.
| Claude Agent SDK | Export a runnable TypeScript agent project. | export |
| LangGraph | Export a Python project — or run hosted on the Calafia runtime. | export + run |
| CrewAI | Export a Python project — or run hosted on the Calafia runtime. | export + run |
How to choose
If you want the agent to keep running here without you maintaining it, take a hosted option and skip this decision — it is the one axis you can defer indefinitely.
Export when you want the code: to read it, to run it in your own infrastructure, or to build on it past what a builder can express. Choose the language you would actually maintain, which in practice decides this more often than any framework feature does.
One thing to be honest about: exporting is a fork. The project is a snapshot, it does not update, and improvements made here after you export do not reach it. That's the trade for owning it.
There is no automated experiment runner yet. Nothing here sweeps this axis for you — changing it means editing the agent and reading the next run. A compare mode that runs one task across several compositions is named on the index as not built, and that is still true.