← Notes

What delegating to an agent does not delegate

Public guidance for building with models is to start from the simplest solution and only add complexity when needed; whoever delegates the building still owns that criterion.

When an agent writes the code, authorship of the implementation changes hands. The criterion of necessity does not.

Anthropic's engineering guidance for people building with models is to "find the simplest solution possible, and only increasing complexity when needed". It is a rule about the decision, not about who types. An agent that picks a retrieval pipeline without anyone having measured the need for it did not break the rule on its own: it worked in a vacuum where the criterion was never stated.

The practical consequence for whoever delegates is that review has to reach the structural choices, not only observable behaviour. A chat that answers is a chat that passed the obvious acceptance test; it may have passed carrying a whole layer nobody asked for, nobody measured, and that now has to be maintained.

The question to ask of every piece of infrastructure inherited from an agent session is the one you would ask a colleague: what measurement justified this. When the answer is none, the piece is not necessarily wrong — it is merely unproven, and a candidate to be simplified before it is optimised.