Tag
Fibr AI's founder shares an experiment testing how AI agents like ChatGPT, Gemini, and Claude perceive their website, finding most sites are built only for humans, and launches a product plus a free checker tool to diagnose AI readability of web pages.
Inherit-MAS 通过在工作流和执行两个层面引入显式继承机制，实现多智能体系统在测试时的演化优化，在 WorkBench 和 HotpotQA 上超越了 EvoAgent、EvoMAS、TacoMAS 等基线，并显著降低了 token 消耗。
FloWright proposes a hierarchical, structure-aware reward paradigm enabling LLM agents in multi-agent workflows to self-evolve and co-evolve without extra models or labels, while DataWright adaptively hardens datasets into workflow-level tasks. Small open models trained with FloWright gain up to +7.41% performance, with co-evolving multiple roles (+5.03%) outperforming optimizing a single role (+2.83%).
This paper conducts the first study of behavioral cost inefficiencies in coding agents, identifying three key behaviors and evaluating mitigation strategies, finding that developer-designed skills can reduce costs by up to 41.73%.
Cursor has reduced token costs in their AI coding assistant by 7% through optimizations like tighter prompts and better caching, emphasizing the value of optimizing for reading in agent systems.
Salesforce AI research shows that when tuning an agent around a weaker model, copying a stronger model can degrade performance; instead, targeted corrections to the weaker model's failures improve results without disrupting the existing setup.
A tweet sharing three interesting papers from Hugging Face Daily Paper on advancements in KV cache compression for local LLMs and agent harness optimization.
Google and Google DeepMind released a paper on Dream-RSI, which introduces a method for recursive self-improvement by having AI agents optimize their exploration strategies using a replay simulator, leading to efficient performance gains in tasks like GPU kernel optimization.
The paper proposes Ecdysis, a framework for training runtime harnesses for LLM agents that identifies recurring failure patterns to improve efficiency and accuracy, achieving 1.84x faster training and 18.56% higher reasoning accuracy.
The article explains why AI startups often have poorly structured prompts with contradictions and ambiguity, and proposes a modular, code-like approach to improve agent quality, reduce regressions, and lower costs.
This paper introduces Speculative Macro Commit (SMC), a two-tier agent system that reduces latency in tool-using LLM agents by speculatively executing future action chains and committing them when matched, achieving speedups over sequential execution.
AutoSaddler is a method for automatically patching agent harnesses from execution traces, ensuring updates generalize by testing on a held-out set to prevent regressions.
MEGA is a self-evolving infrastructure for coding-agent optimization that distills reusable wisdom from sessions, composes it via a typed Wisdom Graph, and uses operational evidence to continuously improve both the agents and the knowledge guiding their optimization.
Kubit launches product analytics aimed at optimizing AI agent actions by analyzing user behavior.
This paper introduces a two-phase continual-learning evaluation on Terminal-Bench 2.0 to test whether gains from agent-optimization methods compound when applied recursively. It finds that only RELAI-VCL, which incorporates regression control, achieves compounded improvement.
A detailed report on optimizing a production vLLM serving configuration on NVIDIA's DGX Spark, correcting flags that were costing 34% MTP acceptance after reviewing 90+ official NVIDIA documents and running a 69-scenario tool evaluation.
Introduces STRACE, a framework that performs structural trajectory analysis and causal extraction to construct high signal-to-noise optimization contexts for improving long-horizon agents, outperforming baselines on a formal verification task.
Bayesian-Agent presents a framework that treats reusable skills and SOPs as hypotheses, using Bayesian inference to guide agent behavior and improve task performance through posterior-guided harness optimization. It achieves significant improvements on multiple benchmarks with deepseek-v4-flash.
Hermes Agent has added native support for skill bundles, allowing multiple skills to be triggered together. The author advises bundling only logically chained workflows to avoid conflicting instructions.
The article highlights how verbose CLI output wastes tokens for LLM coding agents and introduces a pattern-based compressor that reduces shell command output noise while preserving essential information.