A task may need the same operation several times: clean an identifier, compare a record, or extract a particular detail. The model can write a Python helper, use it, and adjust it when the inputs change.
VARIANT-1 also supports optional changes to executable tool methods within a chat. Both forms of adaptation change how work is carried out; neither trains the model’s weights.
Start with a useful helper
A small function gives a repeated operation a name and clear inputs. The model can combine it with other code and tools, while keeping the intermediate data available in Python.
Ordinary helpers work without mutation enabled. They are often enough for a task that needs a repeatable cleaning, calculation, or comparison step.
The helper should be checked against representative inputs. A function that handles one example may still fail on an empty value or an unexpected format.
When a tool method changes
With mutation enabled, a model can propose and activate a revised tool method within its chat. The change has executable source and a recorded state, so it can be distinguished from a suggestion that was never activated.
Mutation is off by default for new chats. Turning it off prevents new proposals and activations; it does not remove methods already active in the chat. Reset and rollback are separate actions.
A better method earns its place in the result.
Try it on different inputs
Reuse is where a method’s assumptions become visible. Try a changed schema, a missing field, or a second set of records. Check whether the output is still correct and whether the method is easier to work with.
Include the effort spent writing, testing, and repairing it. An activated method is not proof of a completed task or an improvement over the previous approach.
Code and live data can separate
A helper’s source may be saved while the data or connection it used is gone. After a runtime restart, check those dependencies before using the method again.
Portable checkpoints can restore supported values when enabled, but do not recreate every live resource. The state guide explains what to expect.
Keep the work open to correction
You can inspect results, steer the task, or request a stop. Stopping does not undo actions already completed in files or other applications.
Mutation does not sandbox Python: code runs with the app’s available Windows permissions. A revised method needs the same clear task scope and output checks as any other code.
The aim is to let methods develop with the work, while keeping their behaviour and results available for review.