What's this about?
A paper arguing that for agent systems, memory and cost are not implementation details to optimise later but architectural decisions that should be made explicitly up front.
What exactly is "context management"?
Deciding, on every step, what goes into the model's limited window: what to keep, what to summarise, what to retrieve, what to drop. Do it badly and the agent forgets the crucial constraint while remembering a stale error message. The paper's point is that this is not tuning — it determines both what your agent can do and what it costs, so it belongs in the design rather than in a later optimisation pass.
What's the community saying?
Practitioners largely agree, with the recurring note that context management is where most real agent engineering time actually goes, and that it is under-discussed relative to model choice — which is the same conclusion the harness papers above reach from the research side.