⍝ APLflow

Write the spec in a notation.
Let the machine write the code.

AI agents made writing code nearly free. That moved the bottleneck to verifying it. APLflow is a working pattern for closing that gap: the human writes intent in a sixty-year-old executable notation, an agent expands it into production code, and the notation itself runs as the independent test oracle.

1 · You · the spec
One line of array notation. It is not pseudocode; it runs. 
avg ← +⌿÷≢          ⍝ mean: sum down the columns, divide by the count
avg 3 1 4 1 5
2.8
2 · The agent · the expansion
A coding agent translates the spec into the language your production stack actually uses. 
def avg(x: np.ndarray) -> np.ndarray:
    """Mean over the first axis."""
    return x.mean(axis=0)
3 · The interpreter · the judge
The notation executes as ground truth. Agreement across random inputs is evidence; disagreement is a caught bug. 
for x in random_arrays(1000):
    assert apl("avg", x) ≈ py.avg(x)
1000 trials · 0 disagreements

The test is independent. When one agent writes the code and another agent writes the tests from the same English prompt, both can share the same misunderstanding and happily confirm each other's bug. A spec you wrote yourself, in a notation with exactly one meaning, breaks that circle.

Ken Iverson designed this notation as a tool of thought in 1962, four years before anyone could execute it. The expensive part was always translating notation into software. That part is now automated. The essay makes the full argument; the loop walks a real example end to end, including the bug it catches.