A long task and a long active context are not necessarily the same requirement. The paper You don't need long context investigates frequent compaction as a way to continue extended work without carrying the entire interaction history into every generation.
Why context length affects cost
The study explains that token generation becomes more expensive as the existing context grows. Its measurements of frontier open-weight deployments show substantial differences in latency, throughput and memory use across context sizes.
The reported disparity can reach seven times the generation cost in the studied configurations. That upper bound is not a universal price multiplier: the models, precisions, workload and hardware assumptions determine the measured result.
Keeping a smaller working state
The authors report little to no task-accuracy loss across a range of long-input and long-horizon tasks when contexts are kept short through frequent compaction. The important qualification is that compaction preserves a useful state; it is not equivalent to deleting history without deciding what information must survive.
For an application, that state may need to retain requirements, unresolved questions and verified progress. Whether a particular compaction policy keeps the right information is an evaluation question, not something the title alone establishes.
One action per context
The paper also reports that performance does not collapse on evaluated software-engineering tasks when each context permits at most one action. From that observation it outlines semantic state space models as a possible direction for agent design.
What to compare in practice
A useful trial would measure completed-task quality, total serving cost and failures caused by lost information, using the same tasks for long-context and compacted agents. This work questions the necessity of large active contexts in its tested settings. It does not establish that every document, dependency or exact quotation can safely be compressed away.
Comments