Tag
This survey provides a systematic framework for self-improving autonomous agents, categorizing approaches into foundation model improvement and scaffolding improvement, and reviews applications and evaluation methods.
The article discusses principles for designing LLM harnesses that are intuitive, transparent, and lean, drawing inspiration from Unix philosophy to reduce cognitive load and improve reliability.
Evaluates agentic LLM systems for generating breast cancer treatment recommendations using 72 clinical cases, finding that the best system (Claude Opus 4.8 with D&C+SA pipeline) achieved a global score of 0.594 but remains insufficient for unsupervised clinical use due to persistent errors.
Presents Oat, a lightweight unsupervised method for identifying error steps in LLM-based agentic failure trajectories using neural controlled differential equations trained only on successful trajectories. It achieves 200-5000x speedup over prompting baselines with significant F1 improvements in in-domain and out-of-distribution settings.
This paper re-evaluates the methodology of automatic harness evolution for LLM agents, highlighting that its gains may stem from additional test-time search rather than improved harness design, and that evaluation on the same benchmark risks overfitting. Experiments show that harness evolution does not consistently outperform simpler test-time scaling methods.
This paper introduces Who&When Pro, a large-scale benchmark for automated failure attribution in AI agentic systems, with 12,326 failure trajectories across 26 benchmarks, and provides insights into how LLMs attribute failures.
The paper proposes Epistemic State Replication (ESR), a belief-replication layer for agentic distributed systems that shifts replication from data visibility to knowledge visibility, allowing semantically equivalent decisions despite divergent generative model outputs.
Anthropic and Google released three free workshops on building self-improving agentic systems, covering multi-agent pitfalls, skill gaps, and building autonomous agents.
Andrej Karpathy outlines five key shifts—memory, multimodality, reasoning, search, and tools—that transform LLMs from simple chatbots into full agentic systems, providing a 40-minute roadmap for next-generation AI workflows.
This replication study independently validates the Natural Language Tools (NLT) framework across 14 models and 8,560 trials, showing that NLT improves tool-calling accuracy by 14.9 percentage points and reduces critical errors by 93% compared to structured tool calling, with the benefits most pronounced for smaller and reasoning models.
This paper presents SCION, an agentic scientific operating system that integrates AI tools for scientific discovery through a Research Execution Plan (REP) and hierarchical multi-agent execution. It demonstrates applications in materials analysis, molecule design, and protein screening, outperforming existing autonomous research-agent baselines.
The article explores whether in agentic systems the causal chain of the process is as valuable as the outcome, and questions if memory architectures should store causal threads instead of just raw outputs.
The author announces a preprint on building an agentic system with composable domains, a verification ratchet, and careful tool naming, sharing practical lessons and an open-source Common Lisp implementation.
A CIO shares observations on the challenges of applying semantic layers and ontologies to enterprise AI agents, highlighting the need for dynamic context selection, continuous evolution of institutional knowledge, and representing judgment rather than just static facts.
This paper presents ElephantAgent, a protocol that enforces Contextual State Continuity to defend against contextual state poisoning in agentic systems, using replicated trusted hardware and historical traceability.
Introduces Janus, a playground system for implementing and evaluating user-involved agentic permission management designs, demonstrating the importance of user input and AI augmentation for privacy and security.
This paper argues that agents should help users construct preferences rather than assuming well-formed ones, proposing the CoPref model and CoShop benchmark. Evaluations show even frontier models achieve only 56% accuracy due to poor preference expansion.
Amazon Web Services launches a $1 billion internal organization for forward-deployed engineers to embed with enterprise customers for deploying AI agents, following similar FDE initiatives by OpenAI and Anthropic.
Anthropic's career article outlines a portfolio roadmap for engineers, covering Python, APIs, RAG, agents, and deployment, with a focus on agentic system patterns like prompt chaining, routing, and parallelization.
This paper examines how LLM-based compression of financial documents can distort investment decisions by losing contextual qualifiers and introducing model-dependent biases, proposing Agentic Context Compression to audit disagreements against the original source.