I have added a design document here to document our choices of interface and implementation. This issue can be used for discussions on the docucment.

The first step is to define how a model will be specified to the engine. To get the ball rolling, here is how pyMC3 defines a model and runs inference:

with pm.Model() as model:
    mu = pm.Normal('mu', mu=0, sd=1)
    obs = pm.Normal('observations', mu=mu, sd=0.1, observed=y)
    trace = sample(2000, cores=2)  # Automatically runs the inference using NUTS