The agent problem is self-management: when to act, what to remember, what to ask

Acting at the wrong time

A proactive agent faces a decision that a reactive one never faces: whether to interrupt at all. Act too eagerly and it spams its user with false alarms; act too cautiously and it misses the moment it was supposed to help. In a preprint, a framework called PRISM is reported to reduce false alarms by 22.78% compared with strong baselines, and to improve F1 score by 20.14% on ProactiveBench, a benchmark for proactive behaviour. m:…1897

The mechanism reported for PRISM is training that gives dense, executable supervision on unlabelled interaction traces, using gate-aligned, schema-locked distillation. In plain terms, the system is taught not just what a good action looks like, but when the gate for acting should open at all. This suggests proactivity is a calibration problem: the agent needs a principled reason to stay quiet as much as a reason to speak. m:…1897

Budgeting attention over a long run

Long tasks have a second self-management problem: attention is a budget, and a fixed workflow spends it whether or not each step still matters. In a preprint, a system called CORAL is reported to reduce input token consumption by 6 to 7 times compared with standard ReAct agents. That is not a small efficiency tweak; it is the difference between a task that stays affordable over many steps and one whose context bill compounds until it collapses. m:…5235

Crucially, the reported saving does not come with a reported collapse in quality. In a preprint, CORAL is reported to achieve 42.8% average accuracy on the GAIA benchmark using an 8B model, and to improve performance on the BrowseComp benchmark by 8.0 points over a strong proprietary agent. The implication is that much of what a long-horizon agent reads is not earning its keep, and that choosing what to carry forward is itself a capability. m:…5235

Remembering skills without re-reading them

Memory is the third budget. An agent that must re-read a full skill description every time it needs a skill pays that cost on every task, and a library of skills quickly becomes a tax on every run. In a preprint, SkillMem is reported to reduce total storage by more than 38 times for a library of 100 skills compared with alternative skill representations, and to achieve up to a 4.23 times end-to-end inference speedup in latent multi-agent reasoning. m:…6612

The reported design compresses skills into a form a reader can consult by role: a role-conditioned cross-attention reader is reported to allow the same latent memory to guide different agents according to their roles and current hidden states. And in a preprint, behavioural distillation is reported to effectively match the change in token predictions induced by a textual skill while keeping the backbone model frozen. The practical read is that skills should be stored once, cheaply, and consulted in a form shaped to whoever is asking. m:…6612

Finding the missing question

The fourth preprint tackles the moment a multi-agent system realises it does not know something. A fixed team of agents, each with a fixed job, will press on with whatever it has. In a preprint, a framework called SPLICE is reported to improve web tool use performance by 8.0% and reasoning and planning benchmark performance by 5.7% relative to static multi-agent skeletons, by locating and closing information gaps instead of following a static plan. m:…4046

That adaptivity is reported to cost something: the SPLICE framework is reported to incur a computational cost increase of 1.25 to 1.44 times compared with static workflows, while its selector mechanism is reported to demonstrate zero-shot transferability across four unseen language-model backbones. Taken together, these four preprints describe the same job from four sides: an agent that can decide when to act, what to keep, what to remember and what to go and find out. m:…4046