Literate Agentic Programming

A codebase can carry its own documentation inside it, written for the artificial-intelligence agent that will later maintain the code. This is a controlled experiment in whether that makes the agent cheaper, faster, and more reliable. The documentation layer under test is called LAP, short for Literate/Anchored Programming.

Cost in tokens and reliability are the two questions the experiment will settle. Elapsed time is recorded, but it is explored rather than settled.

No confirmatory result has been collected. The recorded runs belong to the work of building and testing the measuring instrument. They remain visible because an instrument failure changes what later evidence can mean.

Explore the design Read the proposal Open the Proposal Guide

How the experiment is built: one codebase, three documentation arms, and an open confirmatory question. An animated diagram of the experiment's design rather than a result. On the left, one codebase carries a documentation layer written for a machine to read. In the centre, the same maintenance task is handed to an artificial-intelligence agent working in three versions of that codebase, called arms. The first, labelled LAP, short for Literate/Anchored Programming, carries the full machine-readable documentation. The second, labelled STR, has the same runnable code with that documentation mechanically removed. The third, labelled NON, is an independent unguided control built from the specification alone. On the right, three meters represent the token cost accumulated by each arm beneath an open question mark. Their movement is illustrative and does not encode a measured difference. The confirmatory analysis will test reliability and token cost. Elapsed time remains exploratory. LAP full documentation STR documentation stripped NON no documentation LAP STR NON ? One codebase The same tasks, three documentation arms Confirmatory result remains open