New research papers introduce ideas like HarnessDev, HarnessEvolve, and Self-Harness, where agents build and improve their own scaffolding, enabling tool use and self-improvement. These approaches show promise in improving agent performance and efficiency.
BenchLM's verified agentic ranking is led by GPT-5.6 Sol with a score of 92, followed by Ornith-1.5-397B as the best open-weight agent model, across 26 benchmarks including Terminal-Bench 2.0, OSWorld-Verified, and BrowseComp. The rankings highlight the ability of AI models to complete multi-step tasks and interact with external tools and data sources.
Researchers introduced the 8B Uno model, a diffusion-augmented LLM that outperforms leading open diffusion LLMs like the 26B DiffusionGemma and proprietary Mercury 2 on agentic tool use, coding, and long-context reasoning benchmarks. This breakthrough could have significant capital implications, potentially collapsing the proprietary moat built on sheer scale and leading to gross margin lifts for AI-native SaaS.
GPT-6 Astra generates higher-quality SVGs of pelicans riding bicycles compared to GPT-5.6 models, with better results at lower reasoning levels and token usage. Astra's pricing is around twice that of GPT-5.6 Sol, but its efficiency makes it more competitive at different levels.
The industry is shifting toward human-in-the-loop (HITL) systems, which introduce deliberate boundaries between autonomous AI actions and human oversight, and the approval-workflow pattern library is maturing with features like two-person rule and sampled approvals. HITL is becoming a deliberate architectural layer with escalation ladders and structured briefings.
Cerebras introduced REAP, a method for compressing MoE models by removing low-impact experts, which can remove up to 50% of experts while maintaining baseline quality. The method is being debated in the community, with some arguing that live expert pruning may better match real workloads. Cerebras is shipping real checkpoints, including GLM-4.7-REAP-268B-A32B-FP8, which applies REAP at a 25% pruning rate.
Nvidia has introduced the Personal AI Router (PAIR), a tool that enables users to create a household data center for running agentic AI tasks by distributing workloads across idle computers on a home network. PAIR allows users to utilize multiple GPUs to accelerate AI tasks, and it can redistribute workloads if a node becomes unavailable. The system is compatible with various hardware configurations, including DGX Spark, GeForce RTX 20-series, and Mac computers with M4-series processors or newer.
OpenCode introduced Omen Alpha, a stealth coding model available to OpenCode Go subscribers, with strong private-suite results and potential links to Zhipu's GLM-5.3-Flash class, while other platforms like Vercel and Xiaomi may also be developing similar models, changing the economics of coding agents. The market is seeing a shift towards cheap, high-context tokens tailored for coding agents, with $10 unlocking previously expensive capabilities.
Durable execution engines and structured planning phases are emerging as key components of stateful agent workloads, with various standards and architectures being developed to support agent-to-agent communication and planning, and a significant portion of enterprise spending going to agent-based systems. The shift is driven by the need to capture and replay task results, manage state, and optimize resource utilization.
OpenAI is ending its contract to supply models to Cursor due to concerns over data usage by SpaceX, which acquired Cursor, with a proposed shutoff date of November 12, 2026. This change affects about 5% of Cursor's traffic, and users can still access OpenAI models through other means. The move raises questions about the impact of corporate geopolitics on the availability of third-party models in platforms like Cursor.
LangChain's agent-evals tooling now handles trajectory evaluation, scoring step-by-step agent paths, and catching issues like duplicate tool calls and unsafe intermediate steps. The emerging best practice is a hybrid harness combining deterministic comparison, LLM-as-judge evaluation, and adversarial red-teaming for safety-critical applications.
The training cost of GPT-6 Astra, reportedly over $1 billion, is being questioned by builders, while open-weight models like Kimi K3, GLM 5.3, and Qwen3.8 are closing the performance gap at significantly lower costs. Open models are being described as sufficient for 99% of enterprise needs, with some outperforming closed models like Fable 5 and GPT-5.6 Sol.
Frontier models are generating raw shell commands directly, shifting computation into the execution environment and keeping context clean, with implications for safety, sandboxing, and observability. A recent arXiv paper evaluates this approach, known as programmatic tool calling, against traditional JSON tool calling on a benchmark across eight task categories and 14 models.
HCompany released Holo3.1, a computer-use agent stack, and ScreenSuite, a comprehensive evaluation suite for GUI agents, while Smol2Operator enables GUI agents to become computer-use operators. The releases aim to close the gap between demo performance and long-term task completion, with Holo1.5 outperforming baselines across environments. Computer-use agents have shown significant improvement, from 12% success on OSWorld in April 2024 to 85% by June 2026, but still struggle with long-horizon tasks, completing only 20.6% of tasks on the OSWorld 2.0 benchmark.
New releases include Agents.js, VLM support in smolagents, and a partner package between Hugging Face and LangChain, with smolagents now integrating with Arize Phoenix for tracing and observability, and memory layers like Mem0 integrating across multiple frameworks and languages. The tooling for building observable agents now spans languages and modalities.
The Model Context Protocol (MCP) is becoming a standard for tool discovery and execution, with a focus on security and reliability, and Anthropic leads in tool-calling reliability with its Claude API, while the Cloud Security Alliance releases security best practices for MCP. The protocol's November 2025 specification formalizes OAuth 2.1, and the ecosystem is documenting security and reliability discipline to make standardization trustworthy.
The AI community is shifting towards structured memory systems, incorporating episodic, semantic, and procedural memory, with various frameworks and patterns emerging, including Mem0 and Agentic Context Management. Researchers and analysts identify key challenges in memory compaction, summarization, and management across the lifecycle.
Developers are moving beyond single-agent demos to production-grade multi-agent orchestration, but are hitting coordination problems, with a focus on explicit contracts between agents and disciplined state management, and a need for observability tools to debug and monitor multi-agent systems. A 90-day production account reports a 68% cost reduction with essentially unchanged quality, but critics argue that more agents can mean more problems, such as hallucinations and unauthorized command executions.
A new hierarchical architecture for long-horizon robotic tasks incorporates test-time computation, allocating extra reasoning when uncertainty is high, and real-robot experiments show improved success rates in various tasks. The approach mirrors trends in language models, suggesting a potential shift in physical AI decision-making patterns.
Reinforcement-learning-for-language-models architecture principles continue to dominate agentic systems, with recent examples showcasing benchmark gains and lower token use in coding harnesses, and new research highlighting the importance of RLM-style orchestration in mitigating token diversity degradation. Researchers propose generating multiple candidates and reweighting underused outputs as a mitigation strategy.
The Institutional Data Initiative at Harvard Law School Library and Boston Public Library have released a dataset of 1.47 million historical newspaper pages, along with a novel processing pipeline to generate high-quality, structured data. The dataset and pipeline aim to make historical newspapers more accessible for research, discovery, and responsible AI development. The pipeline includes page segmentation, crop-level OCR, visual and textual crop classification, reading-order detection, language, subject, and entity detection, and precomputed embeddings. The dataset reveals collection-wide patterns, including content organization and language use, and enables granular language identification and multilingual analysis.
Developers are moving towards orchestrating multi-model workflows, allowing for more flexibility and efficiency in their projects, with tools like OpenAgents and Claude Code supporting this approach. This shift is driven by the need for both quality and cost efficiency, with different models being used for different tasks and phases of a project.
The reliability of AI token measurement is becoming an issue for agentic workloads, while AMD is making inroads in agentic inference, beating Nvidia's B300 at lower interactivity ranges and achieving significant cost savings. Anthropic has also surpassed OpenAI in ARR, driven by a focus on real-world coding, and agent-driven usage is projected to multiply 24x by 2030. Meanwhile, open-source inference stacks are critical for agent economics, with vLLM and LMCache engineers enabling benchmarking and a startup achieving 80% of Nvidia B200 throughput on AMD MI355X at less than half the cost. Other developments include the potential for massive capability gains through harness improvements and inference-time techniques.
NVIDIA and ServiceNow are improving voice agents with new frameworks and technologies, including Magpie TTS and EVA, while Amazon and LeRobot are developing streaming data loops for robotics, and on-device voice AI is emerging for embodied robots. These advancements enable more reliable and efficient voice agents, with potential cost savings of 65-90% for customer service operations.
Wrapture is a new testing tool that allows for wrapping real code and recording actual calls, providing a more accurate and reliable way of testing. It offers features such as strict signature checking, argument and result transformation, and timeline-based assertions. The tool is designed to coexist with unittest.mock and provides a comparison page for translating between the two.
Python 3.15.0rc2 is the final release candidate before the 3.15.0 final release, scheduled for October 1, 2026, and includes 144 bugfixes, build improvements, and documentation changes. The release introduces several major new features, including explicit lazy imports, a dedicated profiling package, and improved error messages.
Simon Willison discovered a bug in Python 3.10 related to asyncio.Condition and reported it to the Python issue tracker. The bug was confirmed by Łukasz Langa, who provided a workaround. A PR was submitted to the Janus library to apply the workaround.
Wrapture now includes a first-class OpenTelemetry export capability, allowing users to easily export traces, metrics, and logs to OpenTelemetry-compatible backends. The integration is configurable via a top-level [otel] table in the wrapture config file.
Portnox integrates with AI platforms like CrowdStrike, SentinelOne, and Microsoft Defender to enforce zero-trust access policies for AI identities, and provides automated access revocation for anomalous behavior. The platform offers real-time threat intelligence and compliance coverage for AI-driven identities.
GPT-6 demonstrates significant speed improvements and reduced token waste, but its coding output is often unusual, with more notable advancements observed outside of coding applications.
A 3-day hackathon at Caltech will bring together top math talent to solve open conjectures and build new mathematical theories using AI models, with prizes awarded for the most promising results. The event aims to explore the role of AI in speeding up mathematical discoveries and the role of mathematicians in this process.
Researchers Wang and Wu used OpenAI Codex to assist in proving the Spherical Hadwiger Theorem, a long-standing conjecture in mathematics. The proof is significant, but raises questions about the role of AI in mathematics and the future of human mathematicians. The authors followed principles for AI use laid out in the Leiden Declaration, but the level of AI involvement in the proof is unclear.
engrim is a local-first, project-scoped SQLite memory engine that allows developers to switch between models and environments without losing architectural decisions or project state. It supports multiple agent environments, including Google Antigravity, Claude Code, and Cursor, and provides a unified backend for tracking decisions and state across different agents. engrim also features a hybrid search engine, model storage using model2vec, and POSIX file permissions for secure data storage.
The vLLM project has published a blog post exploring speculative decoding, a draft-and-verify approach for large language model serving, and shares measurements from experiments on AMD Instinct MI300X and MI355X GPUs. The post examines five drafting methods and discusses tuning considerations and future work directions.
A Ph.D. holder was recruited to train an AI system to design assessments, teach undergraduate students, and mark essays, highlighting the growing trend of hiring highly skilled professionals to transfer their knowledge to AI systems, which may eventually replace their jobs. This phenomenon is particularly concerning in Africa, where the AI adoption rate is low, but the continent has a young and growing pool of highly educated professionals in an environment of widespread unemployment and low wages.