NVIDIA researchers report that compacting agent context often matches full-history baselines on long tasks while cutting median peak context from 32K to 8K tokens.
Why it matters: The paper says later tokens can cost up to 7x more to generate on frontier open-weight models, so aggressive compaction could cut costs for teams running long agent sessions.
Brian Armstrong says he has reorganised his workday around a hierarchy of always-on AI agents that feed him one queue of decisions, with humans still approving every change.
Why it matters: Leaders at companies like Coinbase are showing that agents can handle triage and routing for senior decision-makers. That raises the question of where human approval has to stay in the loop.
Researchers from UIC, McGill and other institutions propose Skill++, which rewrites a new skill's metadata so agent routers keep choosing the right existing skills.
Why it matters: Companies maintaining growing skill libraries may see routing errors rise as libraries expand, and poorly written skill descriptions can make interference worse.
Alibaba Research introduces Recursa, which builds a validated skill repository from task experience so a model can adapt to a new domain with fixed weights.
Why it matters: Teams adapting agents to specialised domains may be able to try a validated skill repository before investing in fine-tuning.
FAIR at Meta proposes Routed Continual Learning, which groups past agent runs into clusters with their own memory systems and cuts estimated API cost by 65% to 72%.
Why it matters: A single shared memory for many agent tasks may add little while raising costs. Teams running varied agent workloads may gain from routing memory by task type.
IBM researchers introduce JevAct, a runtime gate between an agent and its tools that checks write calls and hand-offs while leaving the model and tools unchanged.
Why it matters: Teams deploying tool-using agents can catch invalid actions and premature task endings through a harness layer, without retraining the model.
Alibaba Group and South China University of Technology propose SPAR, a self-evolving sales agent that refines its playbook without updating model weights.
Why it matters: Agents that learn from production outcomes can record the wrong lesson when they misread why a result happened. Attribution before memory writes matters beyond sales.
RSI-Dojo is a set of executable environments built from published research, and fine-tuning Qwen models on its trajectories gains 8.74 and 16.87 PostTrainBench points.
Why it matters: The work moves the ability to train models into the weights of a mid-sized open model. Converting research into environments gives a repeatable source of training tasks.
OpenAI introduces LifeVal, a set of 169 agent tasks based on 148 life tasks ranked by 3,837 US consumers, where the most common failure is a wrong answer or recommendation.
Why it matters: Personal assistants depend on recalling past conversations, and the reported win rate is lower there than for saved memories, which affects rollout decisions for memory-based products.
Artificial Analysis has released version 2 of its Decision Index, with 45% of the score now built on private datasets and a headline metric that rewards calibrated confidence.
Why it matters: Teams choosing decision models can see that cost per task and calibration can change rankings, and that per-question billing can make one model several times more expensive than another with similar list prices.