News without the noise
LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning
299 articles · page 7 of 7
The authors present Sakana Fugu, a family of orchestrator language models that generate dynamic, query-adaptive agentic scaffolds to harness the specialized capabilities of multiple LLMs. Fugu and Fugu-Ultra achieve state-of-the-art results on six challenging benchmarks, including SWE-Bench Pro, Terminal Bench, and GPQA-Diamond, surpassing any single publicly accessible model.
Kim et al. introduce Multi-Agent Transactive Memory (MATM), a framework for storing and retrieving agent-generated trajectories in a shared repository accessible to a population of LLM agents. In ALFWorld and WebArena environments, retrieving prior trajectories improves downstream task success rates and requires fewer interaction steps without any coordination or joint training.
The authors investigate whether heterogeneous LLMs can be aligned to directly transfer latent representations (KV-caches) between them, bypassing costly text decoding and re-encoding. They propose a method using a cross-model transformation and two-phase training (reconstruction then generation), achieving performance comparable to or better than text communication in context-aware settings at 2-3x lower compute, and remaining effective in context-unaware settings where prior heterogeneous methods collapsed.
The authors propose Decentralized Language Models (DeLM), a multi-agent framework that replaces a central controller with a shared context and task queue, allowing agents to asynchronously claim subtasks and build on verified progress. On SWE-bench Verified, DeLM improved performance by up to 10.5 percentage points and reduced costs by roughly 50%. On LongBench-v2, it achieved the highest average accuracy across four frontier model families.
The authors test four context engineering strategies for GPT-5 agents using Model Context Protocol tools on a 50-task hotel expense benchmark. They find that pruning context to the last five tool call/response pairs and adding automated summarization yields the best results: 91.6% complete itemization, 99.64% amount itemized, with 63% fewer tokens and 60% less runtime than retaining full conversation history.
The authors present IEA, a conversational image editing agent that manipulates 16 parameterized tools step by step instead of generating pixels directly. Trained via a three-stage multitask pipeline, IEA achieves lower pixel distance on editing tasks and higher ROUGE-L on intent summaries than strong baselines, and outperforms both tool-calling and generative methods in user studies.
The authors propose Astra, an agentic spatial reasoning framework that combines two components: a world simulator (Astra-WM) that generates novel-view images from context images and natural-language camera motions, and an RL-trained VLM policy (Astra-VL) that decides when to query the simulator. Trained with view consistency tuning and a two-phase RL curriculum, Astra improves spatial reasoning benchmarks by 9–10 points over direct VLM answering, demonstrating that effective world-model-augmented reasoning requires learning when, where, and how to imagine.
The authors present the first systems-level characterization of LLM agent memory systems. They introduce a taxonomy, build a phase-aware profiler, and evaluate ten systems across two benchmarks, showing how design choices shift computational costs between memory construction and retrieval. Key recommendations include construction scheduling and amortization via query volume.
The authors propose MAGE, a memory system that treats an agent's history as an execution-state tree rather than a semantically organized document store. On the MemoryArena benchmark, MAGE raised average task success by 7.8–20.4 percentage points over RAG-based baselines while cutting token use by 55.1%.
The authors introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes an AI agent's harness of skills, tools, and workflows using only past task trajectories. By selecting a diverse coreset of challenging tasks, re-solving them in parallel, and using self-validation to pick the best harness update, the method improves pass rate on SWE-Bench Pro from 59% to 78% in a single optimization round.
The authors critique the dominant on-the-fly paradigm for AI agents, which synthesizes plans and executes actions rapidly in response to prompts, arguing it bypasses established software engineering processes like testing and adversarial evaluation. They propose an AI Workflow Store of hardened, reusable workflows to amortize the cost of rigor across users, aiming to improve reliability and security for high-stakes applications.