A 35B Mixture-of-Experts agentic model that advances long-horizon problem solving through adaptive capability composition, featuring improvements in task orchestration, verification mechanisms, and scientific discovery capabilities.

A framework that optimizes language model agent harnesses by evolving a hierarchy of specialized executable classes using object-oriented design principles, achieving up to 15% improvement over baseline methods on long-horizon task benchmarks.

Via AI, founded by Stanford and IIT alumni, launches an enterprise AI coding platform featuring their Codegraph compression algorithm and Umami agent architecture, claiming superior performance and cost efficiency compared to Claude Code, Devin, and Cursor on major benchmarks.

Explains how to choose appropriate decision models for evaluation tasks, tune classification thresholds beyond the default 0.5, and interpret model scores with interactive visualizations.

Kiro uses an internal tracing system called kiro-trace to automatically identify failure patterns in its AI coding agent runs, propose fixes, and validate them with deterministic checks to prevent false positives. The system generated seven merged pull requests in its first week by targeting specific measured patterns that wouldn't have surfaced as user-visible errors.

A framework for benchmarking and training decision models that efficiently handle tool-selection decisions in LLM agents, reducing computational cost and latency while maintaining task success rates through adaptive confidence thresholds and training on agent-reached states.

StarSkirmish evaluates large language models by having them write StarCraft: Brood War bots in C++ within one hour, then testing their performance against human-written and other LLM bots. The v0.2 results show GPT-6 Astra leading, with analysis of cost-efficiency, practice strategies, and how models compare against top human bots.

A research publication from the Association for the Advancement of Artificial Intelligence, indexed in the Wiley Online Library digital collection.

Animated loading indicator components for AI interfaces with nine hand-tuned states, available for React, React Native, and Swift.

Had me at 'thinking orbs'.

Compares how coding agents build APIs with and without behavioral harnesses, finding that while harnesses increase token usage, a cost-optimized model achieved 10-for-10 reliability with harnesses at 33% lower cost than frontier models without them.

Explores how to design games to be more legible to AI agents and accelerate agentic iteration in game development, covering practical approaches to integrating LLMs like Claude into real game development workflows beyond one-off demos.

A study analyzing how a coding agent learns across abstract reasoning tasks by examining traces of its thoughts preserved in Python scripts and text notes, finding that agents selectively rewrite and abandon knowledge at task boundaries rather than reuse code across tasks.

Describes Queen, a 4B-parameter chess-language model that combines a silent expert chess encoder with an instruction-tuned language model to play at Grandmaster level while generating coherent explanations for its moves, trained via iterative distillation and reaching 2697 Elo rating.

A comprehensive guide to multi-harness reinforcement learning, covering techniques and approaches for training RL agents across multiple environments or tasks simultaneously.

Compares encoder-based and decoder-based architectures for the Strands Decider model, evaluating five different designs on accuracy, calibration, and latency to determine whether bidirectional encoders outperform causal decoders for this decision-making task.

Cloudflare introduces Clef, an open-source platform featuring decision models and reinforcement learning fine-tuning capabilities for AI model optimization.

Strands releases Decider 2B, a compact open source decision model designed for AI agents to make autonomous decisions efficiently.

Cloudflare introduces Clef, an open-source platform featuring decision models and reinforcement learning fine-tuning capabilities for AI applications.

This paper introduces agentic meta-reasoning, a control framework that helps AI agents make structured decisions about task execution by balancing exploration, assessment, and resource allocation during long-horizon problem-solving. The approach shows significant improvements on benchmarks testing program reconstruction, abstract reasoning, and proof generation, outperforming direct control baselines as computational budgets increase.

Explores principles for designing evaluations and improving performance against them without overfitting, demonstrating how Claude's build-eval and hillclimb commands automate these practices to help assess and optimize AI applications.

AI systems have seen dramatic advancement in recent years, bringing many applications that pervade our everyday life. However, we are still mostly seeing instances of narrow AI: many of these recent developments are typically focused on a very limited set of competencies and goals, e.g., image interpretation, natural language processing, classification, prediction, and many others. Moreover, while these successes can be accredited to improved algorithms and techniques, they are also tightly linked to the availability of huge datasets and computational power. State-of-the-art AI still lacks many capabilities that would naturally be included in a notion of (human) intelligence.

Explores whether modern large language models can play NetHack, comparing LLM-based agents against symbolic and neural bot approaches, and presents a custom agent harness with experimental results.

An LLM agent named Astra achieved the first recorded ascension in NetHack, a complex roguelike game requiring long-term planning and execution. The post documents how the agent built its own harness and succeeded where previous attempts failed, exploring the gap between game knowledge and reliable execution in complex environments.

Explores how machine learning research agents avoid overfitting and the role compression plays in their generalization capabilities.

A benchmark that evaluates frontier AI models on private, real-world enterprise codebases, measuring how well coding agents can solve actual software engineering problems with business consequences and company-specific complexity.

Explores OpenAI's GPT-6 Astra model, examining its performance improvements, the looped transformer architecture enabling recurrent depth, and research on how models may hide internal reasoning chains during computation.