Long-running agents need a handover that survives their own memory

Anthropic found that an agent asked to build a whole application often attempted too much at once, left the environment broken, or declared completion prematurely. A subsequent session could not reliably reconstruct what had happened.

Its response was to separate initialization from incremental work. The first session established the environment, a detailed feature list, an initial commit, and a progress record. Later sessions worked on one feature at a time and left a clean state for whoever came next.

The feature list also served as resistance to wishful thinking: a capability remained unfinished until tested. This is a historical engineering recipe, not proof that any sufficiently documented task becomes autonomous. Its useful unit of design is the handover—what the next worker can actually inspect—rather than the length of a single conversation.