Long-horizon tasks

Long-horizon tasks refer to AI agent workloads that require sustained, multi-step operation over extended time periods — as opposed to single-turn or short exchanges — and present distinct architectural, infrastructure, and memory challenges beyond raw model capability.

Core characterization

Lance Martin (Anthropic) identifies long-horizon tasks as a central target for agent engineering, but notes that the standard loop-based architecture "is kind of tricky for long horizon agents," implying that naive loop designs do not straightforwardly scale to these workloads. →

Enabling primitives

Martin argues that the paradigm of loops — particularly verifier loops — paired with high-capacity frontier models constitutes "a very good general primitive for long-running asynchronous work." → → This framing positions the loop-plus-frontier-model combination as the foundational building block, even while acknowledging its limitations.

What needs to come together

Martin emphasizes that model capability alone is insufficient. "To build real agents that can operate in these long time horizons, a bunch of things need to come together" — specifically: architecture, infrastructure, security, and memory improvements must converge. →

Memory: the dreaming mechanism

One concrete technique Martin highlights is dreaming — described as "an offline or out of band process that we've used to consolidate and improve memory." This positions memory consolidation as a component specifically designed to support the extended temporal scope of long-horizon tasks, drawing an analogy to how biological memory is processed during sleep. →

Summary of dependencies

Requirement	Role in long-horizon tasks
Loop architecture + frontier models	General primitive for async execution
Verifier loops	Quality control across steps
Memory / dreaming	Consolidation across long spans
Infrastructure & security	Operational stability at scale