Search: “dair-ai--x”

Nice little insights on doing autoresearch with coding agents. Hand a coding agent a dataset, an eval script, one editable file, and no supervision. That's autoresearch and it tries to optimize the nu

Very cool paper from Microsoft. The idea is to train agents on replayed teacher trajectories instead of live environment rollouts. On-policy distillation for agentic tasks is expensive because every u

New research from NVIDIA. Does AdamW have a scale ceiling? This work claims yes, and shows where it sits. At batch sizes up to 100M tokens for next-token prediction, SOAP and Muon maintain training st

The Top AI Papers of the Week (July 20 - July 26): - GAMUT - PRO-LONG - Harness Handbook - From Memory to Skills - Progressive Disclosure - Global Workspace in LLMs - Structured Output Collapses Diver

Diffusion LLMs can now handle real agentic work. LLaDA 2.2 is the first large-scale diffusion LLM built to operate as a real agent, planning, calling tools, and self-correcting across long multi-turn

// Agentic Context Management // Great read for the weekend. (bookmark it) Production agents fail less on reasoning and more on what sits in their context. Conversation history, big prompts, huge tool

New research with Microsoft's and colleagues on training agents inside the harnesses they actually run in. (bookmark it) Why it matters: Agents today live inside elaborate harnesses like Claude Code,

Dynamic workflows are a generalization of harnesses, automations, loops, routing, and graphs. It's the most powerful feature I have built into my agent orchestrator. Supports all kinds of patterns tha

Great paper on self-improving agent harnesses. (bookmark it) If you maintain a production agent harness, finding every file behind one behavior is often harder than writing the edit. Harness Handbook

// Programmatic Memory Enables Long-Horizon Reasoning // Keep the entire interaction log and search it. It works great and beats bespoke memory harnesses on long-horizon tasks. New research introduces

Are structured outputs in agents always good? This paper suggests that you might have to take a closer look. Your product's structured output surface is measurably more homogeneous than the chat surfa

I agree with what this AI paper suggests. Self-improving agents should evolve their benchmarks too. (bookmark it) Self-improving agents are one of the most important directions in AI right now, and mo

New research from Meta. (bookmark it) Most factuality work checks whether the claims in an answer are correct. GAMUT goes after the harder question of whether the answer covers everything it should. I

Progressive disclosure in agents doesn't scale. And its benefits seems agent harness dependent. (bookmark this one) Finally there is a proper study on using agent skills and the effect of progressive

Not because Andrej is saying it but I think voice is goated. And you can mix it with other modalities for even richer prompting. I recorded a session a few weeks back to demo the power of multimodal p

This is a neat feature. I wrote an article a few weeks back about how I built this into my agent orchestrator: https://x.com/omarsar0/status/2073404610501329247?s=20 But I made it multimodal from the

Very cool idea to convert memory to skills. (bookmark it) Most agent memory systems retrieve past traces as passive context. MSCE turns them into executable skills instead. The training-free framework

// Global Workspace in LLMs // arXiv paper for the popular J-space work from Anthropic. (bookmark it) The short recap: If you build on chain-of-thought or steering vectors, this work provides a mechan

Highly recommended. I've often claimed there's huge alpha in building agent harnesses. Turns out harnesses are compositional generalizers. The RLM harness is an instance of this. This could lead to in

Huge launch from @tryramp. Different steps in an agent workflow can use different models. This can help reduce costs significantly without sacrificing performance. Model routing will become a core par

Interesting finding on frontier models. It turns out that frontier models can write proofs but stumble on faithfully copying a long block of text. This has huge implications. It sounds trivial, which

Never too late for any company to get in on open models. Inkling is a MoE transformer with 975B (41B active). Native multimodal support. Context lengths: 64K and 256K. Will be testing and providing th

NEW paper from Microsoft and colleagues. Debugging agent trajectories at scale is challenging. This is a clever approach to monitor and improve agents in production. The problem: Finding which step in

Highly-recommended overview of metacognition in LLMs. (bookmark it) Interesting behaviors in LLMs like confidence calibration, self-verification, knowing when to stop, and knowing what you do not know

Huge if true! We are talking about a 27B multimodal model that runs locally on a phone. That's wild! Bonsai 27B reaches up to 163 tok/s in 1-bit and 134 tok/s in Ternary on an NVIDIA GeForce RTX 5090.

New research from Google DeepMind on effective model routing. LLM routers get judged on accuracy and cost. Both can look great while the router is meaningless. If every model in your society responds

You can now use Claude inside After Effects. Higgsfield's new MCP connector lets Claude work inside your actual AE project. It can build compositions, set keyframes, write expressions, and run the rep

// An Anatomy of CLI Coding Agent Trajectories // (bookmark it) When your coding agent fails a task, when did the run actually go wrong? Most reliability studies use the final label to answer this. Th

Everyone keeps asking me how to build a second brain or an LLM wiki. Here is the easiest setup I have found. I took my Wiki Builder skill, installed it into HyperAgent (@hyperagentapp) as a reusable s

Another big reason to use combination of frontier models. Chain-of-thought monitoring is treated as a reliable safety layer for agents. This DeepMind-affiliated study shows the layer can be argued out

Cool open-source release. LingBot-World 2.0 holds 720p at 60 fps in real time and stays coherent for a full hour of interaction. A 1.3B variant runs on a single consumer GPU, so you can actually run a

This post highlights a practical AI agent that operates intuitively without requiring explanation of its underlying mechanics, suggesting it has achieved a user-friendly interface or autonomous functi

New research from Meta. (bookmark it) It's on how to fix agents that forget previously made decisions. It's well know that long-horizon agents keep forgetting decisions they already made. Meta researc

This post discusses positive experiences with GPT-5.6's capabilities in verification, advisory functions, and high-level orchestration tasks, suggesting the model performs better than expected in thes

A Visual Introduction to Information Theory (bookmark it) Information Theory is such an beautiful and powerful subject. In the era of AI, it's worth spending time learning about it. Here is a highly-r

Here is my other heavily used pattern. Evaluator/Judge: Fable 5 Executor: GPT-5.5 I no longer wait for frontier models or am loyal to any. I now spend more time on better orchestration, harness, skill

Is Meta AI back? Haven't seen Mark post in three years here. Plus, the model is available via API. Not to mention the courage to announce it the same week as the long-awaited GPT-5.6. Great timing if

Muse Spark 1.1 available in new Meta Model API. Somewhere near Opus-4.8/GPT-5.5 level. 1M context window! Computer-use capabilities sound great: Write scripts when automation is faster, click when dir

// The Harness Effect // (bookmark it) Now more that ever pay very close attention to the orchestration harness and its effect on costs and performance. This study ran 22 evaluation tasks on six found

Claude power users: 'Fable 5 is the best' Codex power users: 'GPT-5.6 is the best' Reality: Loyalty to a single model provider is a terrible strategy. The smart choice: clever orchestration between fr

Great writeup from the University of Oxford. It's a taxonomy of LLM-based agent limitations. Good read for anyone shipping with agents. Benchmark scores keep climbing, yet the same agent failures resu

Highly-recommended read. Aligns with what I see in my own harness: > Pi harness got the same success rate as harnesses from the LLM vendors with Opus and GPT, but at 2x less cost > GLM 5.2 was a major

I am more excited to try this pattern instead. Same Executor-Advisor setup but with GPT-5.6 as the executor and Fable 5 as the advisor. It already works wonderfully using GPT-5.5, so I think 5.6 shoul

Massive open-source release! LingBot-Video isn't about video quality; it optimizes for physical reasoning, and it runs sparse: only 3B of 30B params are active at inference. MoE is finally showing up

// Memory becomes an action space // Great paper on long-term memory for agents. (bookmark it) In short, it's discusses the use of a learned policy for using memory at the right granularity. Most memo

Nice stats on usage of open models across OpenCode. GLM-5.2 is still underrated, but one of the models that has really surprised me on agentic tasks is deepseek-v4-flash. Extremely cheap and effective