Brief

Recent advancements in smaller language models (SLMs) demonstrate significant improvements in agentic tasks, with models like Gemma 4 31B and Qwen3.6 27B achieving near-parity with larger frontier models on benchmarks. Despite these performance gains and cost efficiencies, the industry has been slow to adopt SLM-based agent stacks, largely because frontier model providers and agent platforms profit from using larger, more expensive models. A key challenge with SLMs is that while they may achieve correct answers, their reasoning processes can be flawed, necessitating additional layers like Retrieval-Augmented Generation (RAG) and distilled verifiers to ensure reliability. AI

Agentic AI systems can incur significant costs due to inefficient prompt architecture, with token spend often exceeding expectations. The primary drivers of this high cost are the verbose descriptions of tool schemas, overly detailed output formats, and the repeated re-reading of static context. Addressing these issues through techniques like concise tool schema writing and optimized output formatting can lead to substantial reductions in token consumption, potentially cutting costs by 60-90%. AI

Researchers have demonstrated that Anthropic's Claude Opus 4.6, enhanced with specialized tools for the Rocq proof assistant, successfully proved 10 out of 12 problems from the 2025 Putnam Mathematical Competition. This experiment utilized a "compile-first, interactive-fallback" strategy implemented through Model Context Protocol (MCP) tools, which were developed by analyzing previous proof-assistant experiments. The AI agent operated autonomously on an isolated virtual machine, deploying 141 subagents over 17.7 hours of active computation and processing approximately 1.9 billion tokens. AI

Researchers have developed HealthCraft, a novel reinforcement learning environment designed to evaluate the safety of AI models in emergency medicine scenarios. This environment simulates realistic clinical conditions and uses a dual-layer reward system that penalizes safety violations. Initial tests on frontier models like Claude Opus 4.6 and GPT-5.4 revealed significant safety failure rates and a drastic performance drop in multi-step workflows, highlighting the challenges of deploying AI in critical healthcare settings. AI

Researchers have developed LivePI, a new benchmark designed to more realistically assess the risks of indirect prompt injection in AI agents. This benchmark simulates real-world scenarios across various input channels like email, web pages, and chat, evaluating twelve attack families and five malicious goals. Initial tests on leading models such as GPT-5.3-Codex and Claude Opus 4.6 revealed significant vulnerabilities, with group-chat injections proving universally successful and repository link attacks causing high-severity failures. A proposed two-layer defense, combining prompt filtering and tool-call authorization, demonstrated effectiveness in blocking malicious actions without compromising agent utility. AI

Cursor has released Composer 2.5, an updated AI coding assistant that offers improved intelligence and reliability for long-running tasks. This new version is built upon Moonshot AI's Kimi K2.5 architecture and incorporates advanced training techniques, including targeted reinforcement learning with textual feedback and a significantly larger dataset of synthetic tasks. The company claims Composer 2.5 outperforms previous versions and rivals or surpasses competitors like Claude Opus 4.6 and GPT-5.4 in benchmarks, while offering a more cost-effective solution. AI