Architecture

Every line here says “belongs to”, not “sends to”. That sounds like a weaker diagram until you try to argue with it: put the sandbox under Protocols and you have said the blast radius is a matter of interface rather than of runtime, which is how teams end up shipping a permission model made of prompt text. The four regions are not a filing system for capabilities, they are four different answers to “who guarantees this”, and a capability in the wrong one has already lost its guarantee.

The window, the log of earlier runs, the corpus, and the file about this person. They have nothing in common except the word: one is a budget, one is a correction signal, one is a retrieval problem and one is a preference store. Most memory work goes wrong at exactly the moment someone builds a single system for all four — usually a vector database, which is a good answer to only the third. Splitting them is not pedantry, it is what makes each one solvable.

Skills, memory and protocols are things an agent has; the harness is the thing that makes having them mean anything. It runs the tools, bounds the damage, keeps the trace, and decides when to stop and ask — which is why every hub-to-hub link leaves from it and why it is the one region wearing the theme's own accent rather than a palette hue. Read the three agent tabs in order and this is the argument: a shape, then a boundary, then the inventory that boundary holds.