搜索“Agent”

y members of the Claude Code team use HTML instead of Markdown to produce richer, more readable, and easily shareable outputs. Markdown has become the dominant file format used by agents to communicate with humans. It’s simple, portable, has some rich text capability and is easy to edit. Claude has even gotten surprisingly good at using ASCII to make diagrams inside of Markdown files. But as agents have become more and more powerful, I’ve found that Markdown has become an increasingly restrictiv

匹配字段正文

up on a busy few weeks of open-weight LLM releases. The thing that stood out to me is how much newer architectures are focused on long-context efficiency. As reasoning models and agent workflows keep more tokens around (for longer), KV-cache size, memory traffic, and attention cost quickly become the main constraints, and LLM developers are adding a growing number of architecture tricks to reduce those costs. The main examples I want to look at are KV sharing and per-layer embeddings in Gemma 4

匹配字段正文

minal Bench 2.0: a smaller, sharper context produces less work for the model and a meaningfully smaller bill at the end of the run. What’s driving the token efficiency Most coding agents assemble context through grep and keyword search. While this approach has improved in quality over time, it remains inefficient: agents burn turns crawling files, reading large spans of code, and pulling in irrelevant matches just to find the few lines that actually matter. Every miss costs another round trip, a

匹配字段正文

ssion to save bandwidth during transfer and avoid bottlenecking the network during decompression. But we also try to keep image creation fast, especially to better support dynamic agent workloads. You can fast-forward through tens of seconds of application host-side startup with CPU memory snapshotting. Now, let’s consider our third step: start the application program on the host, ready it to service requests (tens of seconds) This encompasses all of the work that needs to be done to go from the

匹配字段正文

e gap to hardware efficiency. Together is continuing V4 kernel work across the prefill and decode paths. Developers should benchmark V4 in their actual regime. A 1M-context coding agent and a short-context chat assistant exercise different parts of the serving stack. V4 widens the gap between workload shapes. Long-context agents benefit most directly from the architecture. They read large amounts of cache during decode, so V4's compressed KV layout, larger tensor-parallel configurations, batchin

匹配字段正文

y shape their unit economics. Estimates put inference at 80-90% of the total lifetime cost of a production AI system , simply because it runs continuously across every user query, agent step, and API call. And while training is a bounded investment, inference scales with every new user and use case you ship. At NVIDIA GTC 2026, NVIDIA CEO Jensen Huang framed this shift plainly: “People pay for information, but people mostly pay for work. Agentic systems get work done.” That shift from AI as a no

匹配字段正文

to most effectively structure your prompt, use tools, and layer on compaction. It is often said in engineering that "cache rules everything around me", and the same rule holds for agents. Long running agentic products like Claude Code are made feasible by prompt caching which allows us to reuse computation from previous roundtrips and significantly decrease latency and cost. At Claude Code, we build our entire harness around prompt caching. A high prompt cache hit rate decreases costs and helps

匹配字段正文

Learn how the Claude Code team designs, tests, and evolves tools by thinking from the model's point of view. One of the hardest parts about building an agent harness is constructing its tools. So how do you design your agents' tools? Do you give it one general-purpose tool like bash or code execution? Or fifty specialized tools, one for each use case? To put yourself in the mind of the model, imagine being given a difficult math problem. What tools would you want in order to solve it? It would d

匹配字段正文

d designs, and getting it to build complete applications without human intervention. This work originated with earlier efforts on our frontend design skill and long-running coding agent harness , where my colleagues and I were able to improve Claude’s performance well above baseline through prompt engineering and harness design—but both eventually hit ceilings. To break through, I sought out novel AI engineering approaches that held across two quite different domains, one defined by subjective t

匹配字段正文

d architectures are still a novelty, and the main selling point is mainly (long-context) efficiency versus just modeling performance. Hence, I think they are a great candidate for agent contexts (like OpenClaw). Personally, I think the problem with hybrid architectures is also that the inference stacks are not quite as optimized, yet, and I find that I get better tok/sec throughput when running LLMs locally using more classic setups like GPT-OSS with grouped-query attention. Anyways, I am curiou

匹配字段正文

man deliverables, which often leave much more informative first impressions). The rest of the post remains the same as it was on March 3 2026. Summary: Opus 4.6 can, with a simple agent scaffold, create mostly-playable but somewhat broken CLI versions of Slay the Spire and Balatro1. Intro Last weekend I was trying to think of really difficult tasks we could give to AI agents to upper-bound their capabilities. I thought of two examples: Recreating a basic version of the video game Slay the Spire

匹配字段正文

sign, but requires you to build inside its component framework. We needed determinism from the outside, applied to arbitrary web content. Second, our videos are generated by an AI agent. Constraining the agent to Remotion's component model would mean teaching it one library's idioms instead of letting it use the entire web platform. The less framework surface area the agent has to reason about, the better the output. So: no special framework. No library buy-in. Just a URL. This meant building th

匹配字段正文

never released as an open-weight model.) The obligatory benchmark overview shows that Qwen3.5 exceeds the previous Qwen3-Max model across the board, with a much stronger focus on agentic terminal coding applications (the main theme this year). Qwen3.5 appears to be roughly on par with GLM-5 and MiniMax M2.5 in terms of pure agentic coding performance (e.g., SWE-Bench Verified). Since the Qwen team likes to release a separate coding model (e.g., see Qwen3-Coder-Next, which we discussed previousl

匹配字段正文

ywhere. After configuring, make sure you select Finished to save your desired configuration. Selecting a model Ollama shows a model selector with recommended and available models. Agents work best with at least 64k context length. Ollama’s cloud models have full context length, which provide the best experience with agents like OpenClaw. Cloud models: Local models (requires GPU VRAM): Running securely OpenClaw has the ability to read files and execute actions when tools are enabled. Ensure that

匹配字段正文

Ramp uses Modal to power Ramp Inspect, an internal background coding agent that now writes over half of all merged pull requests at Ramp. Leveraging Modal Sandboxes, Ramp spins up full development environments in seconds, giving every builder at the company access to what is effectively unlimited parallel engineering capacity grounded by Ramp’s internal codebase, enterprise context, and developer tooling. The idea: A coding agent as fast as local, accessible to everyone Ramp's AI team saw an opp

匹配字段正文

ago, for the vast majority of people, “using AI” meant talking to a chatbot in a back-and-forth conversation. But over the past few months, it has become practical to use AI as an agent: you can assign them to a task and they do them, using tools as appropriate. Because of this change, you have to consider three things when deciding what AI to use: Models, Apps, and Harnesses. The exact same model, Claude Opus 4.6, asked the exact same question, “Compare ChatGPT and Claude and Gemini” in three d

匹配字段正文

Written by Nicholas Carlini, a researcher on our Safeguards team. I've been experimenting with a new approach to supervising language models that we’re calling "agent teams." With agent teams, multiple Claude instances work in parallel on a shared codebase without active human intervention. This approach dramatically expands the scope of what's achievable with LLM agents. To stress test it, I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kern

匹配字段正文

aluations 100 hard browser tasks, one leaderboard; The judge; Results; Why Cloud scores higher; Using the benchmark 100 hard browser tasks, one leaderboard To truly understand our agent performance, we built a suite of internal tools for evaluating our agent in a standardized and repeatable way so we can compare versions and models and continuously improve. We take evaluations seriously. As of now, we have over 600,000 tasks run in testing. This is our first open source benchmark. BU Bench V1: 1

匹配字段正文

Introduction Good evaluations help teams ship AI agents more confidently. Without them, it’s easy to get stuck in reactive loops—catching issues only in production, where fixing one failure creates others. Evals make problems and behavioral changes visible before they affect users, and their value compounds over the lifecycle of an agent. As we described in Building effective agents , agents operate over many turns: calling tools, modifying state, and adapting based on intermediate results. Thes

匹配字段正文

h V2, it achieves gold-level performance on math benchmarks. However, the model is also trained with tool-use in mind and also performs well on other tasks, for instance, code and agentic tasks. At the same time, the DeepSeek team writes about computational efficiency as a big, motivating factor. That’s why they use the Multi-Head Latent Attention (MLA) mechanism from V2 and V3 together with the DeepSeek Sparse Attention (DSA) mechanism, which they added in V3.2. In fact, the paper says that “De

匹配字段正文