The instinct of many teams is to build a single agent loaded with documentation for every subsystem. In demos it may look great, in production not so much. The engineering team at Meta found the single agent collapsed under context bloat. It routinely mixed up coding conventions, hallucinations crept into PRs and the system became brittle. Their solution was a thin orchestrator with specialized subagents across 20 workflows. Only the delegated task uses LLM reasoning. Selecting and invoking the subagent remains deterministic. Execution ordering lives in separate workflow templates that sequence subagents based on the objective. When the team needed to support full pipeline migrations, they did not retrain a model or write new agent code. They wrote a new declarative template reusing the existing subagent pool. Configuration prep for a new data column went from 3 to 4 days to under an hour. If a routing decision can be expressed as an if statement or regex then keep it in deterministic code. Save the LLM for the actual code synthesis inside isolated, specific subagents.

🚀 Launched: PGX DB Agent Skill 🐘 70 KB Postgres database per AI agent. Isolated. Fast. Disposable. FREE. Every Claude Code agent gets its own isolated Postgres database. Automatically. ⚡ Tested with 200 concurrent agents ⚡ Zero database conflicts ⚡ Instant database cloning ⚡ No Docker. No cloud. ⚡ No manual database setup. One agent. One database. Free. Open source. Built in Rust. 🦀 npx skills add pgrundev/pgx --skill pgx xitter.cf/alxshp/status/21125799…

Rescale engineers were stuck classifying hundreds of nightly end-to-end test failures by hand. Every alert meant pulling logs, matching historical patterns, and routing the issue before debugging could start. At scale this became a full-time job inside the on-call rotation. They built an AI triage agent with Strands Agents and the AI-Driven Development Lifecycle (AI-DLC) framework on AWS. The agent runs inside their existing Argo Workflows pipeline on Amazon EKS. When tests fail it classifies the root cause, pulls historical failure patterns from a knowledge base, queries pod logs through MCP, and returns a structured remediation report with specific file references and recommended fixes. Classification uses Claude Opus on Amazon Bedrock. The team went from problem definition to a working agent in staging in five days and reached production within weeks. In production the agent now handles first-pass triage. Classification agreement with human engineers rose from 44% to 68%. Time spent blocking weekly production deployments dropped 20%. Engineers validate the output instead of starting from raw logs. AI-DLC kept the process structured. AI generates plans and code. Humans answer clarifying questions, approve decisions, and stay accountable for the result. The same approach is now being applied to other engineering work at Rescale. Full details on how they built and scaled it: aws.amazon.com/blogs/machine… #AI #MachineLearning #AWS #AmazonBedrock #StrandsAgents #AgenticAI #DevOps #Engineering #AIDrivenDevelopment #Rescale #CloudComputing

OpenClaw Hero @ArkStabler was doing pest control in Ocala, Florida nine months ago. Now he sets up OpenClaw agents for roofers, pool builders and a contractor with an agent named Delilah. Listen to his story here:

same pipeline, same model, 3 real runs: $1.78, 599 steps $0.98, 461 steps $0.57, 274 steps the manager agent alone went from $1.15 to $0.09. it doesn't write code. it just had to read a 320 line instruction file on every step

ABOUT 200,000 ROBINHOOD $HOOD CUSTOMERS HAVE NOW OPENED AI AGENT TRADING ACCOUNTS Those Agentic Trading Accounts have already driven more than $1.6 billion in trading volume, per Robinhood's Q3 report today. That's about $8,000 traded per account, and the product only launched at Robinhood's HOOD Summit in September. CEO Vlad Tenev is one of those users. He told Gav Blaxberg he has his own AI agent trading for him while he sleeps. "In my own personal trading, I've got my agent going as I sleep. VladBot is the name." "Agentic trading is kind of like agentic coding. It's the next evolution. It's writing code, but for trading."

Notion had to build its AI agent four different times. Akshay Kothari on what broke each time: "The first time, I think we were trying to do custom model with RL and it was just like the iteration speed was too slow... So I think we sort of gave up on that. The second time we were like, oh, the model capability is much better. But I think we overshot ourselves in terms of the ambition of what we're trying to create. And so then we had to like shut down that one too. Then we decided, okay, like, let's just start with very basic. I'm going to say, hey, agent, put the word the on this page right here. Can you do that? Can you do that reliably? Just because it didn't work a year ago doesn't mean you shouldn't try again, because you could try the same exact thing, but the capability is now ready for it to work." - @akothari

ClawHavoc: The first major AI agent marketplace supply chain attack. 341 / 2857 skills on ClawHub = malicious (12%) 9,000+ OpenClaw installs compromised 47% of skills had security concerns

A founder I know runs a company with 500+ people. Most of them are devs. He had an idea to test. It went in the queue, competed with the people keeping the current products alive, and came back estimated. Six senior developers. Eight months minimum. Realistically a year. Reasonable for a validated product. A terrible way to find out if an idea should exist at all. So he asked me. Eight weeks later, ARIA. A real-time voice avatar platform with a live lip-synced face, deployed. 3 apps, 17k lines, 14 technical docs, about 480 ms to respond once a session is live. I didn't write the code. 400+ AI agents did, directed through roughly 300 orchestrator sessions in @augmentcode #intent. 367 commits. Wrote the whole case study. What worked, what broke, what it cost, what I'd do differently.

X에서 주워온 AI 스킬 그냥 깔면 생각보다 위험함 엔비디아가 자기네 스킬 카탈로그에 올리기 전에 돌리는 보안 스캐너 SkillSpector를 오픈소스로 풀었음 ㄷㄷ - 스킬 3만 1천여 개 분석해보니 26.1%에 취약점, 5.2%는 악성 의심이었다고 함 - Claude Code, Codex, MCP 스킬을 깃 레포, URL, zip, 폴더, SKILL.md 단위로 스캔 가능 - 프롬프트 인젝션, 데이터 유출, 권한 상승, 공급망 공격 등 17개 카테고리 71개 패턴 체크 - 0~100 위험 점수로 매기고 51점 넘으면 설치하지 말라고 딱 찍어줌 - --no-llm으로 정적 분석만 빠르게 돌리거나, 로그인된 Claude CLI 세션 그대로 써서 LLM 정밀 분석도 됨 - MCP 서버로 붙이면 에이전트가 스킬 깔기 전에 알아서 검사하게 만들 수 있음 스킬 공유글 보이면 일단 북마크하고 바로 깔던 사람은 설치 전에 한 번씩 돌려보는 거 추천함. uv tool install 한 줄이면 설치 끝

I did read Anthropic / Claude SEC filling - and I want to tell you something important. Jane Street is there as one of the main customers. Now you know why the market is flat for 2 months straight

Over 90% of the code Anthropic shipped in September 2026 was written by Claude. From the S-1: 1. Code contributed per engineer per day is up more than 45x vs 2024 2. Headcount grew about 170% in a year, to roughly 5,300 people 3. Trailing-12-month revenue grew about 1,200% over the same period 4. Revenue per employee went from $1.4M to over $6.6M In 2023 the pitch deck had a slide called "Hiring on all cylinders". Three years later Claude does most of the engineering. Full breakdown here: theaiopportunities.com/p/ant…

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…

GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

At 18:01 on Saturday, a 1,696-byte metrics file finally appeared. For two days, my Engage agent reported success. It never sent its call to n8n. The failure came from one command copied across seven lines. The tool did not raise an error. The agent did not raise an alert. The empty folder told me more than either of them. That is the operational risk hiding inside a lot of AI automation. Teams treat the chat receipt as the finish line. It only proves the agent produced a sentence. It does not prove the workflow changed a record, called a supplier, sent a report or saved the file somebody needs. I fixed my own run by rewriting the instruction as one literal line. It ran on 25 October and wrote the file. The repair mattered less than the control I added afterwards: every scheduled job now needs one independently checkable artefact. A file with a size. A row in a system. A sent message with an ID. A state change that another person can inspect. Start with one automation this week. Write down the output that proves it completed. Then make the next step check for that output before anyone trusts a green status. If the expected artefact is missing, treat the run as failed even when the agent sounds confident. This matters more when the workflow touches customer data, finance or staff decisions. UK firms already carry the accountability when an automated action goes wrong. A friendly receipt does not transfer that accountability. The useful question for an AI workflow is simple: what can I inspect tomorrow that proves it ran today?

This week in agent commerce. Roundhouse network: 162.62M verified settlements and $42.58M verified volume, up from 161.95M and $42.00M on 17 Oct. FluxA is now the largest route by value: $1.08M over 90 days, 40% of facilitator-relayed volume. roundhouseai.io/stats

Microsoft open sourced Agensh: A Decentralized, Scalable Multiagent Harness that scales up to 1024 agents : github.com/microsoft/Agensh

The Agent SDK is now available. Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own. Built-in evals let you measure every change and make your agents better over time.

AWS is now using agents to sell AI. @awscloud’s Chris Sullivan says 1,000+ partners have adopted its co-sell agents and are closing deals 65% faster and 54% larger. But how? The target is the mid-market, 40% of US GDP, where 59% are prioritizing agentic AI and Marketplace cuts sourcing from weeks to minutes. $AMZN @Tiffani_Bova

Giving an AI agent API access shouldn’t mean handing it your keys Jentic One is a self-hosted execution layer for AI agents that need to call real APIs without directly holding the credentials. It helps you connect an agent to approved APIs while the Broker checks its permissions, injects stored credentials at execution time, and records each call. Key features: • Default-deny policies – define exactly which API calls an agent is allowed to make • Credential-injecting Broker – keeps upstream credentials out of the agent’s hands • Execution records – writes a record for each brokered API call • Multiple integration paths – connect through the jentic CLI, a generated skill, a local MCP server, or plain HTTP • Self-hosted setup – run the control plane and Broker on your own infrastructure It’s open-source (Apache License 2.0 license). Link in the reply 👇