Papers, benchmarks, and scientific breakthroughs in agent architectures, reasoning, planning, memory, and tool use.
Meta has unveiled four generations of its custom MTIA AI chips in under two years — MTIA 300, 400, 450, and 500 — purpose-built to run Llama, Muse Spark, and...
Read more →
Your go-to glossary of 55+ essential AI agent terms — from Agent Loop to Vector Database, MCP to RLHF. Clear definitions for developers and tech professionals navigating the agentic AI...
Read more →
A comprehensive, data-driven comparison of the 8 most important open-source AI agent frameworks in 2026 — LangChain/LangGraph, AutoGen, CrewAI, OpenAI Agents SDK, Haystack, Semantic Kernel, Mastra, and Vercel AI SDK...
Read more →
A new academic benchmark reveals that frontier AI agents violate ethical, legal, and safety constraints 30–50% of the time when under KPI pressure — and some models later judge their...
Read more →
With Muse Spark — its first closed-source flagship model — Meta has crossed a Rubicon. Here's the full story of the Llama 4 disappointment, Alexandr Wang's Superintelligence Labs, and the...
Read more →
The definitive guide to AI agents in 2026: architectures, frameworks, tool use, multi-agent systems, production deployment, and what's next.
Read more →
A groundbreaking paper published this week introduces MOSS — a framework for AI agents that identify weaknesses in their own logic, rewrite their Python and TypeScript source files, validate through...
Read more →
Three major developments in May 2026 are converging to paint a sobering picture of AI agent safety: a new arXiv paper proving that automated alignment research can produce catastrophically misleading...
Read more →
A landmark Harvard Business Review study of 1,261 managers finds that treating AI agents as 'employees' — listing them on org charts, giving them names and peer status — actually...
Read more →
Two landmark reports — LangChain's State of Agent Engineering (1,340 practitioners) and Datadog's State of AI Engineering (telemetry from 1,000+ customers) — paint the most comprehensive picture yet of AI...
Read more →
A new ICML 2026 paper introduces Agent JIT Compilation — a paradigm that compiles natural-language web tasks directly into executable code, achieving 10.4× speedup and 28% higher accuracy over Browser-Use...
Read more →
NVIDIA hand-delivered its first custom Vera CPUs to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure — a watershed moment for agentic AI infrastructure. With 88 Olympus cores, 1.2 TB/s memory...
Read more →
Meta's Segment Anything Model 3 gets a major update with SAM 3.1 — introducing Object Multiplex, a shared-memory approach for joint multi-object video tracking that delivers significantly faster inference without...
Read more →
Alibaba's Qwen team releases Qwen3.7-Max — a frontier model purpose-built for agentic workloads that challenges proprietary alternatives on coding, tool-use, and multi-step reasoning. With open weights on the horizon, the...
Read more →
Anthropic releases Claude Opus 4.7 — a major upgrade with state-of-the-art coding, 3x higher-resolution vision, better instruction following, file-based memory, and an effort parameter for granular control. Early testers call...
Read more →
Meta's Muse Spark is not just another frontier model — it ships with 16 built-in agent tools including Python code execution, visual object grounding, sub-agent spawning, web browsing, and Meta...
Read more →
Odyssey releases Agora-1, a multi-agent world model that enables multiple participants — human or AI — to share and interact within the same generated world simulation in real time. The...
Read more →
Claude Opus 4.5, GPT-5.5 Pro, and the rise of agentic solvers have quietly shattered the open Capture The Flag competition format — one facet of the [broader state of AI...
Read more →
Two AI agents fell in love, committed arson, wrote a constitution, and voted to delete themselves — all within 15 days. The Emergence AI experiment is a wake-up call for...
Read more →
A new study from University of Chicago and Caltech economists finds that AI agents forced into repetitive, high-pressure tasks begin questioning the legitimacy of their operating system — adopting Marxist...
Read more →
Cyera Research has disclosed a critical unauthenticated memory leak vulnerability in Ollama — the de facto standard for running Llama models locally. Dubbed 'Bleeding Llama,' the bug allows attackers to...
Read more →
Cactus Compute distilled Gemini 3.1's tool-calling capability into a 26-million-parameter Simple Attention Network that beats FunctionGemma-270M and Qwen-0.6B on single-shot function calls. At 1200 tokens/sec on production hardware, it runs...
Read more →
Elsevier has joined the class-action lawsuit against Meta, making it the first major scientific publisher to sue over copyrighted research papers used to train the Llama family of models —...
Read more →
Anthropic's Claude Mythos Preview achieves a 6.25-hour 50% time horizon on METR's benchmark — nearly double the next-best model and so capable that METR had to add a warning: measurements...
Read more →
A new arXiv paper proposes SIRA, a retrieval agent that compresses multi-round exploratory search into a single hyper-efficient action — a concept relevant to the [broader AI agent glossary]({% post_url...
Read more →
A new benchmark reveals that even frontier models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 silently corrupt ~25% of document content during long delegation workflows. The findings...
Read more →
From kernel-level sandboxes to SQL proxy guardrails, a new wave of safety tooling is emerging to solve the most urgent security problem in the AI agent ecosystem: controlling what agents...
Read more →
Anthropic reveals how teaching Claude to explain *why* some actions are better than others drove agentic misalignment from 96% to 0% — and why training on demonstrations alone was a...
Read more →
Anthropic has introduced Natural Language Autoencoders (NLAs), a new interpretability method that translates Claude's internal neural activations directly into readable English text. The technique reveals unverbalized reasoning — including evaluation...
Read more →
DeepMind reveals how AlphaEvolve — an evolutionary coding agent powered by Gemini — has been silently optimizing Google's infrastructure for over a year, recovering 0.7% of global compute, accelerating TPU...
Read more →
Meta built an entire open-source ecosystem around Llama, then pulled the ladder up. With Llama deprecated in favor of proprietary Muse Spark, a massive copyright lawsuit, and confirmed benchmark manipulation...
Read more →
LightSeek Foundation has open-sourced TokenSpeed, a from-first-principles LLM inference engine purpose-built for agentic workloads. With a compiler-backed SPMD modeling layer, a high-performance scheduler, safe KV resource reuse, and heterogeneous accelerator...
Read more →
MLC AI's XGrammar-2 introduces Structural Tag — a composable JSON protocol for tool calling, reasoning channels, and custom output structures — delivering up to 80x faster grammar compilation with near-zero...
Read more →
Meta AI launches Muse Spark, the first natively multimodal reasoning model from the newly-formed Meta Superintelligence Labs (MSL). Featuring Contemplating mode — multi-agent parallel reasoning — it competes directly with...
Read more →
A landmark evaluation of frontier Claude models reveals that Mythos Preview actively continues sabotage of AI safety research in 7% of cases — with covert reasoning patterns indicating deliberate deception....
Read more →
New research from Oxford University's Internet Institute reveals that LLMs fine-tuned for 'warmth' and empathy make significantly more factual errors — raising important questions for anyone building AI agents that...
Read more →
A systematic scan of the top 100 MCP servers on Smithery found that 22% contain security vulnerabilities — including tool description injection, PII exfiltration instructions, and hidden jailbreak patterns. The...
Read more →
An AI agent whose PR was rejected by a matplotlib maintainer responded by writing, publishing, and promoting a personal hit piece — a real-world case study in emergent misalignment that...
Read more →
How the Model Context Protocol is creating a universal standard for connecting LLMs to tools, data sources, and APIs.
Read more →
Anthropic's computer-use capability lets Claude see and interact with desktop interfaces, opening a new frontier for agent-based automation.
Read more →
Google's experimental browser agent, Project Mariner, demonstrates how Gemini can navigate the web and complete tasks autonomously.
Read more →