Change the model. Keep the institution.

What Symphony asks of an AI model is that it can use tools and follow directions. It is not written around one vendor's way of doing things.

So a frontier model, an open model running on hardware you own, or both inside one project are all ordinary. The model is a component you choose, not a company you join.

The claim, stated carefully

We are not telling you model choice does not affect quality. It plainly does, and a smaller model running on your own hardware is not as capable as the largest hosted ones. Our own testing records where they fall short, by name and by size.

What we are telling you is that your work does not start over when the fashionable model changes — and it will change, repeatedly. Your rules, your history and the judgement your team has accumulated are not hostages of one lab's release calendar.

What you give up without it

Compounding. Chase each release and every change becomes a migration: prompts retuned, conventions relearned, and everything the tool had learned about your work left behind with the old one. The speed is real, and it is rented.

Spending you control

The organisation sets its own limits — per agent, and per model. The same model can be used two ways: generously for work that earns it, frugally for work that does not. You get a cost estimate before a request goes out rather than a bill afterwards, and when a request falls back to a different model, you are told which model actually answered instead of being left to assume.

Where this stands today: the per-agent and per-model limits above are in the hub. The desktop application applies one output limit across every model rather than letting you set them individually. We would rather tell you which half you are getting than describe the better half and let you assume.

The unusual part

We keep a scored log of open models on our own hardware, and it records the failures — including one small model marked "not viable" for agent work. It is not published yet. When it is, it goes up with the failures in it, or it does not go up.

We publish it because the results are only useful if they include the ones that went badly. A log of successes tells you what we chose to run twice.