The Agentic Daily

Agents are paying first with tokenized cards, not tokens. Visa, BILL, and Billtrust are piloting AI agents that pay buyer-approved invoices through supplier portals. Per Visa, the agent pays only invoices the buyer has already approved, using tokenized cards. The first agent payments run on rails AP teams already trust, with an approval step they already know. The rail underneath will keep changing. What every AP team will keep asking for is proof of who approved what before it moved. corporate.visa.com/en/sites/…

We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git. Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage.

Developers on X don't even fathom the levels of agentic engineering thats cooking at @openclaw, dropped CI bill by >95%. Was nearly $1m p/m. Beyond all the obvious ones was adaptive testing using tree based classifier and/or system 1 model based on past test behaviour.

Imagine replacing $349/month in SEO tools with just one AI prompt. With AI/ML API MCP + Jev, your AI agent can: → Analyze your website and competitors → Find and score 10K+ keywords → Build and launch Google Ads campaigns All for just $0.61 per campaign. No subscriptions. No extra signups. AI agents are making marketing easier, faster, and cheaper. This is actually impressive.

BREAKING $AMD Q3 2026 Earning Result 🚨🚨🚨 Revenue: $13.38B vs $13.09B est ✅ EPS: $2.09 vs $1.94 est ✅

1/ Made top 6 at the OpenAI DevDay Community Hackathon 🎉 I built strata. Agents love taking shortcuts across your codebase, and enough of them and the whole thing collapses. strata catches them early. Here’s OpenClaw growing to 54,322 files across 524 districts 👇

@Reuters’ latest reporting on Iranian jet fuel deliveries to Myanmar centered on the MARIS (formerly the REEF) (IMO 9263382), which was sanctioned by the United States in 2024 and has been central to the Iran-to-Myanmar jet fuel trade since October 2024. To hide the origin and destination of her cargo from the international community, the MARIS spoofed her AIS (automatic identification system), broadcasting a false path that would suggest to any casual onlooker that she had traveled between Iraq and Bangladesh. In reality, she was repeatedly detected by SynMax’s maritime domain awareness platform, Theia, at the Port of Shahid Rajaee in Bandar Abbas, Iran, and later in Yangon, Myanmar. Theia automatically analyzes more than 40 million km² of medium-resolution, broad-scale satellite imagery daily, fusing it with global AIS, synthetic aperture radar, and other datasets to detect and attribute tens of thousands of vessels, regardless of their AIS output. Because every detection is archived against the vessel’s record, the MARIS’s pattern of life can be reconstructed retrospectively, and any deviation from it stands out. Atlas, SynMax’s Agentic AI, works alongside Theia to interrogate those pre-processed, auditable datasets, highlighting behavior consistent with deception. Where a detection warrants closer examination, high-resolution imagery can be cued over the vessel’s location. On the 12th of September 2026, the MARIS was imaged alongside in Yangon by @Satellogic Newsat-51 (50cm), despite her AIS placing her hundreds of kilometers away. Broad-scale imagery shows what is happening on an oceanic scale. High-resolution imagery confirms it beyond any doubt. We are grateful to Reuters for their continued collaboration on this reporting. Read the full investigations in the comments.

We added a new sign up page for Agents. We're building Coinbase to be the most AI-enabled financial account. Give it a try: coinbase.com/agents

No te puedes perder estas noticias de IA hoy — 3 de noviembre de 2026 Las novedades que vale la pena seguir hoy. Más noticias en Facebook: facebook.com/profile.php?id=…

Applied Compute update: - crossed ⁠$100M annualized revenue run rate up from zero 18 months ago - raised $350M at $3.25B led by Elad Gil - 40% of customers use AC for both custom model training + production inference - inference serving 1T tokens a day after launching 12 weeks ago - 2X enterprise customers since last raise - 3X team since January, 85% are researchers and engineers ⁠- Working with customers like Microsoft, Shopify, Harvey, DoorDash, Cognition, Clay, Decagon and others To move even faster we need people who are extremely technical AND extremely practical. If this is you, please reach out.

The best AI agents fail 88% of real professional work. Nobody tells you that. 📊 SemiAnalysis released a new benchmark this morning. GDPx. 206 real tasks across 11 professional domains, built with 14 specialist data partners. Not abstract puzzles. The work professionals actually do. The scores are brutal. Claude Opus 5.5, the best model tested: 12%. Gemini 4 Carbon: 10.9%. GPT-6 Astra: 9.6%. Read that again. The best score on real professional work is 12 out of 100. Let me be clear about what this is and is not. It is one research firm's benchmark, and it is deliberately hard. It says a lot about what is demos and what is daily work. Every demo you have ever seen is cherry-picked. The one showcase that worked, filmed twice, edited once. GDPx is the opposite: a spread of real tasks, no retakes. Here is what a business owner should actually conclude. The gap is not a reason to skip AI. The gap is a reason to never run it unattended. An agent that gets 12% right needs a person checking the output. That is not a bug in your setup. That is the setup. The business that wins is the one that pairs a capable agent with a sharp reviewer, and lets the trust curve do the rest. Exactly the "review before relax" loop we keep coming back to. And if the best agent scores 12%, the accountant who trusts theirs with 100% of the work is either lying to you or lying to themselves. The demos are 100%. The real work is 12%. Plan accordingly.

Building an AI support agent is easy Building one that doesn’t refund a customer twice is the hard part. Customer says: “My order hasn’t arrived, Refund me” Behind the scenes the agent must: • Understand intent • Authenticate the user • Fetch order + payment status • Check refund eligibility • Call the right tools • Execute the refund • Verify it succeeded • Confirm to the customer This breaks in production when: • The LLM times out • Tool arguments are invalid • Payment succeeds but the response times out • The agent retries and refunds again • A consumer crashes mid-event • The agent loops on tool calls • Someone tries prompt injection • The refund amount exceeds the limit A production agent needs: • Idempotency keys on every money-moving action • Retries with backoff + jitter • Durable workflow state • Tool-level validation and authorization • Circuit breakers • Step and token budgets • Human approval for high-value actions Observability and monitoring: • Trace every step of the agent run (intent, tool calls, decisions) • Log tool inputs and outputs without sensitive data • Track latency, token usage, and retry counts per run • Alert on duplicate refunds, high failure rates, or budget breaches • Record the final outcome and any human interventions • Make it easy to replay a failed run with the same request ID Core rule: Let the LLM decide what should happen. Let deterministic services decide what is allowed to happen. That’s the difference between a chatbot and a system you can trust with money. What’s the riskiest action you’ve seen an agent try to take without proper controls?

AI stats worth watching: 98% of enterprises have pulled back an AI deployment due to failures or unexpected behavior, while 89% redesigned use cases due to rising token costs. The Objective breaks down Cloudera’s latest research: bit.ly/4eC9EVE sprou.tt/1oQOi2Kv3qo

If someone can steer your AI agent, they can steer the MCP servers it calls. AWS disclosed CVE-2026-116635 in the AWS Serverless MCP Server (awslabs.aws-serverless-mcp-server before v0.2.1). Under certain conditions, a path bug lets that actor read JSON files outside the deployment status directory, and create or overwrite them if write access is enabled. Update to v0.2.1 and keep write access off unless you need it.

Interesting. Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM. Decagon reportedly runs ~90% of its workloads on open weight models. Why? Because when you're running AI agents in production for e.g. customer service, cost and latency matter enormously. You don't need a frontier model with PhD-level reasoning or even a model that knows the capital of Finland to handle every customer interaction. We're seeing similar dynamics across hypergrowth AI startups and increasingly the large enterprises we work with. But most importantly, the future isn't binary. When use cases are new, you want the smartest general-purpose model available. But once workflows mature and become predictable, smaller, faster, specialized models increasingly win. I wouldn't be surprised if 80–90% of routine enterprise AI workflows eventually run on open-weight and smaller models. This doesn't mean frontier labs lose. It means the market is splitting into two massive opportunities: frontier intelligence and specialized execution at scale.

OpenClaw 最受欢迎的 5 个 Skills ！！！ 大语言模型有一个共同的问题：对话结束，学到的东西就没了。 OpenClaw 的 Skill 系统解决了这个问题。一个 Markdown 文件就是一个技能，写好之后永久加载，不需要你反复教。 ClawHub 是 OpenClaw 的官方技能市场，排名前五的 Skills 总下载量超过 43 万次。 第一名：Self-Improving Agent。 这不是一个工具，而是一个学习系统。它把 AI 犯的错、你纠正的话、发现的新知识结构化记录下来。 三层进化机制：即时记录（错误日志）→ 知识沉淀（学习日志）→ 永久能力（晋升到 SOUL.md / AGENTS.md）。 最终，成熟的学习可以被提取为全新的 Skill——一个 AI 的进步，变成所有 AI 的进步。 第二名：Tavily Web Search。 专为 AI 设计的搜索引擎。不返回一堆 HTML 链接，而是干净的结构化摘要，附带来源。支持深度研究模式和新闻搜索，让 AI 的知识不再有截止日期。 第三名：Find Skills。 最有意思的一个。它让 AI 自己去发现需要的能力。你说"帮我做 PR Review"，它自动在 ClawHub 搜索，找到合适的 Skill，问你要不要装，装上就直接开始用。AI 自主发现能力、获取能力、使用能力。 第四名：Gog。 一个 CLI 操控 Google 全家桶。Gmail、Calendar、Drive、Sheets、Docs、Contacts。一个命令搜邮件，一个命令查日程，一个命令写表格。你说帮我看上周邮件整理到 Sheet 里，三个服务联动，一张周报出来。 第五名：Summarize。 80 页 PDF、两小时 YouTube 视频、长篇网页。一行命令压缩成结构化摘要。支持多个 AI 模型，支持多种格式。 组合技：Skill 之间的工作流编排 让OpenClaw自动盯盘: Tavily 搜索 → Summarize 摘要 → Gog 写入表格和邮件 → Self-Improving Agent 记录优化 → Find Skills 发现新能力。 单个 Skill 是工具，组合起来就是自动化工作流。 一个耐人寻味的现象： 排第一的不是搜索，不是工具集成，而是自我进化。 社区用下载量投了票。AI 的元能力比任何具体能力都更值钱。 而这一切的起点，就是一个 Markdown 文件。

56% of organizations say their agentic security is established and purpose-built. Yet 90% had an agentic security incident in the past year. Only 36% of those contained it in under 24 hours. So we looked past the self-assessments. We surveyed 3,766 security leaders and scored their organizations on what they’ve actually built: how agents get registered, when security gets involved, what they can see, who owns it, and whether they can prove it on demand. 13% have a defensible posture. Our new report breaks down what they do differently, plus five moves to get there. Read Deployed but Ungoverned: bit.ly/4klcEcD

if your AI agent doesn't have a harness, congratulations: you are the harness every frontier lab is quietly obsessed with harness engineering right now, yet almost nobody explains what it actually does here is the concept explained simply: think of the model as a wild horse. fast, strong, but runs wherever it wants the harness is the reins, the saddle, and the cart. without it, all that raw horsepower goes nowhere useful. claude in your browser and claude code run the exact same underlying model. one chats. the other autonomously inspects repos, runs tests, and ships production code. the only difference between them is the harness. in claude code, the harness is simple: → a persistent rules file (CLAUDE.md or AGENTS.md) → sandboxed access to your terminal and file tree → an execution loop that runs the code, checks the diff, and self-corrects here is why you need one: without a harness, you are the execution loop. you repeat the system prompt every morning. you manually paste error logs. you verify every output. you decide the next step. Google researchers kept the model completely frozen and only changed its harness. on one real office task, the score jumped from 43 to 78 out of 100, just from adding a rule to double-check its own work. three principles of an elite harness: 1. write the rules once a single CLAUDE.md or AGENTS.md file loaded before every task. stop typing "follow our style guide" into chat windows. 2. prune the tool surface give the agent only the tools the specific job needs. every extra tool is an extra surface for hallucinations and context drift. 3. verify before exit unit tests for code, checklists for everything else. if the harness cannot verify the work, you will have to. how to start: take the instruction you paste most often and save it as a persistent rule. take the mistake the model makes most often and turn it into an automated check. frontier models are already commoditized. the harness is where all production value lives.