搜索“Agent”

do tasks that take 100 hours by the end of 2026 ( 448 ). This significant rise in the length of time that AI systems can work independently correlates neatly with the explosion in agentic coding tools - this is the productization of AI systems which do work on behalf of people, acting independently for significant periods of time. It also loops back to AI R&D, where if you look closely at the work of many AI researchers, a lot of their tasks boil down into things that might take a person a few h

匹配字段正文

cessible, current documentation. Stale docs and tangled code aren't just engineer pain points. They're inference inputs; Fast CI/CD pipelines. If your build takes 40 minutes, your agent's feedback loop takes 40 minutes too; Education and time to absorb it. DX's data showed something counterintuitive. Light AI adoption actually decreases productivity. Only moderate-to-heavy adoption outperforms non-adoption. The learning curve is real, and it requires time, safe-to-fail projects, and psychologica

匹配字段正文

Enterprises are converging on cloud agents as the future of software engineering — and many are concluding they should build their own. Posts like Stripe's, detailing how they built a homegrown cloud agent, make the path look achievable. Building cloud agent infrastructure requires two investments: the technical infrastructure to run agents securely and autonomously in the cloud, and the change management to make agents productive across your engineering org. We've spent over two years on both,

匹配字段正文

10 months ago, I wrote Don't Build Multi-Agents, arguing that most people shouldn't try to build multi-agent systems [1]. Parallel agents make implicit choices about style, edge cases, and code patterns. At the time, these decisions often conflicted with each other, leading to fragile products. A lot has changed since then. At Cognition, we've begun to deploy multi-agent systems that actually work in practice. Our original observations still hold today for parallel-writer swarms: most of the sex

匹配字段正文

moments; Try it A SKILL.md and some Python helpers. The LLM has complete freedom. If something's missing, it writes it. The learning A few months ago we wrote The Bitter Lesson of Agent Frameworks. The argument: don't wrap the LLM in abstractions. Maximal action space, then restrict. We were still wrapping its tools. Every click(), type(), scroll() helper is an abstraction you decided the model needs. Every one of them is a constraint the RL'd model has to fight around. Why raw CDP When we built

匹配字段正文

ult/larger tasks, tasks that aren’t straightforward SWE tasks, and tasks that aren’t easy to programmatically check. Also, when I apply AIs to very difficult tasks in long-running agentic scaffolds, it’s quite common for them to reward-hack / cheat (depending on the exact task distribution)—and they don’t make the cheating clear in their outputs. AIs typically don’t flag these cheats when doing further work on the same project and often don’t flag these cheats even when interacting with a user w

匹配字段正文

Agents are becoming extremely effective at finding security vulnerabilities. They are relentless in analyzing code and you can spin up multiple of them to go through source code quickly. given enough agents, all bugs are shallow— Johann Rehberger (@wunderwuzzi23) February 10, 2026 It is an emerging capability that many security researchers and bug bounty hunters have observed over the last year. Gadi Evron posted about the upcoming AI Vulnerability Cataclysm last year to help raise awareness. In

匹配字段正文

same thing, and when the tool is built around one provider, that's basically true. The prompts are tuned for that model. The context retrieval is optimized for its strengths. The agent behavior is shaped around its quirks. Switching the model means rebuilding all of it. That's why it feels like a migration. But those are separate layers, and they don't have to be coupled. Think of it as three things: The model is the LLM doing the generation: Claude, GPT-5, Gemini; The harness is what feeds the

匹配字段正文

" because it extends even up to "applications" (e.g. the Ghostty GUI app has more forks than ever before with customized patches on top, which is awesome). The factory of today is agentic3. I say that as objective truth, regardless of what your feelings about it are. You can argue that 99% of the stuff coming out of these factories is total crap, but you can't argue the sheer quantity of stuff coming out. The numbers are everywhere spanning tech stacks and industries and they're undeniable. AI i

匹配字段正文

fessor and leader of the HeavyBall Research group at NYU Shanghai. One example of applying this academic engine to production use cases involved one of the leading real-time voice agent companies. They came to Together with a hard constraint: time-to-first-64-tokens above roughly 100ms breaks the conversational experience. On their previous setup, deployed on NVIDIA B200 GPUs, they were hitting 281ms. Fast for most workloads. Not fast enough for theirs. Together's kernels team worked with them t

匹配字段正文

Alexis Gallagher & Rens Dimmendaal March 12, 2026 Fans of vibecoding and agentic tools say they are 2x as productive, 10x as productive – maybe 100x as productive! Someone built an entire web browser from scratch. Amazing! So, skeptics reasonably ask, where are all the apps? If AI users are becoming (let’s be conservative) merely 2x more productive, then where do we look to see 2x more software being produced? Such questions all start from the assumption that the world wants more software, so th

匹配字段正文

enClaw captured some of this). A website can surface its capabilities so that a subset can just show up in ChatGPT, be it a real estate search or a shopping cart. You’ll tell your agent to look at a recipe on Instagram and order the ingredients on Instacart. Everything can get piped to everything else, and everything can talk to each other! Meanwhile, (saying the quiet part out loud), if you could set and control those APIs and manage the flows, that gives you power. Standards have been a basic

匹配字段正文

owing the same log-linear improvements[5] that pre-training showed. As frontier labs train on a broader and broader set of RL tasks — first math, then code, then a wide variety of agentic work — the models will generalise. Just as GPT-2 generalised beyond its training distribution when trained on a broad enough internet scrape, RL-trained models will generalise beyond their specific training tasks when trained on a broad enough set of RL environments. At some point, the frontier model clears the

匹配字段正文

e reasons to be skeptical of the magnitude. Early Work on Monitorability Evaluations We show preliminary results on a prototype evaluation that tests monitors' ability to catch AI agents doing side tasks, and AI agents' ability to bypass this monitoring. How Does Time Horizon Vary Across Domains? We build on our time-horizon work and analyze 9 benchmarks for scientific reasoning, math, robotics, computer use, and self-driving in terms of time-horizon trends; we observe generally similar rates of

匹配字段正文

weight model (gpt-oss, and I wrote a standalone article about it earlier this year); MCP (joining the Linux Foundation) has already become the standard for tool and data access in agent-style LLM systems (for now); I expected the ecosystem to remain more fragmented in 2025, until at least 2026 9.2 Predictions for 2026 We will likely see an industry-scale, consumer-facing diffusion model for cheap, reliable, low-latency inference, with Gemini Diffusion probably going first; The open-weight commun

匹配字段正文

hen the material fights back, you’re pushing with that confidence. Think of it in LLM terms. The design work you do before is the goal, context, and prompt you’re building for the agent to address the problem. The execution work is where you guide the agent through specific decisions. After reading all this, you might think, who has time for this? We need to ship. Why not build right away and figure it out as we go? I don’t think it comes down to time. My explanation might have been long, but th

匹配字段正文

ted in different ways: your phone could know what you bought on Amazon or what images you saw in Instagram, but it wouldn’t see the graph that led Meta to show them to you. Now an agentic LLM assistant might become another blind man feeling the elephant, in different ways: it knows different things, but as you use it it also knows different thing about you, drawing new conclusions, especially if it’s buying things on Amazon and Instacart for you, and you’re using its browser, and its wearable. I

匹配字段正文

arized Devin's strengths and weaknesses in real-world environments, with examples and metrics from customers. We hope this will be helpful to anyone who's interested in real-world agent deployment. Strength pattern 1: Junior execution at infinite scale Devin excels at tasks with clear, upfront requirements and verifiable outcomes that would take a junior engineer 4-8 hrs of work. Unlike a human, though, it is infinitely parallelizable and never sleeps. This makes it well-suited to critical but l

匹配字段正文

the right. The AIME is a hard math exam, GPQA tests scientific and legal knowledge, the MMLU is a general knowledge test, SWE-bench and LiveBench test coding, Terminal-Bench tests agentic ability. Data from Epoch AI . Despite these issues, all of these benchmarks, taken together, appear to measure some underlying ability facto r. And higher-quality benchmarks like ARC-AGI and METR Long Tasks show the same upward, even exponential, trend. This matches tests of the real-world impact of AI across i

匹配字段正文

oject, and person. The system uses the past completion history, existing backlog, and future roadmap as context to understand what to do; Simple changes are dispatched to a coding agent and completed without additional intervention. More complex requirements are automatically collated into project specifications and as new requirements roll in, the system keeps documentation up to date with the latest information In this new world you, the user, are still in control—you decide which projects to

匹配字段正文