Future Direction: Horizon Scaling (Session to Scientific Program)

Description

Fourth of six open design directions (Section 12.4). Asks how the architecture's turn/session/sub-agent units support long-horizon dependability as autonomous work extends beyond a single session into multi-session scientific programs, hypothesis generation systems running over days, and algorithmic discovery spanning weeks. Kwa et al.'s METR 50%-time-horizon metric gives an empirical frame. Open: whether the harness layer alone closes the gap, whether a cross-session memory substrate (see Section 12.2) is required, or whether horizon-scale work demands coordination primitives beyond session/sub-agent/memory.

Key claims

Relations

Sources

src-20260423-0cff68d3291b