Our direction, while Bubbling AI is in development. This is a design commitment, not a performance benchmark or a release schedule.
The software around an AI model is sometimes called a harness: the context, tools, checks, and working loop that turn an answer into useful work. We want that surrounding system to work with a model’s strengths and adapt as those strengths change.
A complicated workflow is not automatically a better one. Extra planning, repeated instructions, and unnecessary handoffs can get in the way. We want each part of the design to earn its place by helping the person or enforcing a boundary that matters.
That is the direction we are taking with Bubbling AI. We want progress in models to become progress in your work.
Keep the purpose, revisit the workaround
A design that helps with one generation of models can become unnecessary later. An extra planning step, a rigid template, or repeated instructions may have solved a real limitation. That is a reason to test whether it still helps, rather than preserve it indefinitely.
Our approach is to revisit the experience as models change. Keep the parts that improve quality, clarity, and control. Simplify or remove the parts whose original purpose has disappeared. A more capable model should not have to struggle through yesterday’s assumptions.
That does not mean removing permissions or privacy boundaries because a model appears more trustworthy. Those controls serve the person and the organization using the product. They have a different purpose from a workaround for model behavior.
Keep important boundaries outside the model
We do not assume a model will follow every instruction on every step. Access checks, required approvals, and supported workflow gates belong in the system around it. The model cannot grant itself permission or substitute a claim of compliance for a required check.
Those controls have a lasting purpose. A workaround for a temporary model limitation may not. Improving the harness means knowing which is which and testing changes against the work people actually do.
Choose models for the work
Different tasks place different demands on a model. A quick organizational step, a difficult comparison, and a careful piece of writing do not necessarily call for the same balance of capability, speed, and cost.
We are designing the workflow so different kinds of models can contribute where they are useful. That may include models from commercial providers and models with publicly available weights. The label alone does not decide quality, licensing, privacy, or suitability.
The goal is a useful overall result with understandable constraints. Offering every model at every moment is a different promise, and one we are not making. Available models will depend on evaluation, provider terms, cost, and the data handling required for the task.
Improvement needs evidence
A new release is a reason to evaluate. It is not proof that every task will improve.
We want to look at the things a person experiences: the quality of the discussion and finished work, the effort needed to correct it, response time, reliability, and the cost of getting the result. We also need to check that a change preserves the permissions and data boundaries people rely on.
Sometimes the right improvement will be a stronger model. Sometimes it will be a clearer question, better context, or less unnecessary work around the model. We intend to keep revisiting those choices instead of treating the first design as permanent.
A workspace that keeps developing
Bubbling AI is new. We are building for continuing improvement and intend to explain meaningful changes as they arrive. We are not promising a fixed update frequency, automatic support for every new model, or a performance advantage we have not measured.
The point of keeping the system adaptable is to give people a better partner over time: more capable help, clearer choices, and less friction between an idea and useful work.
Read why context management matters, why human judgment stays central and why approvals belong in the product.