Hugo

@ihteshamali Curious how much of the win came from decomposition vs the model itself.

A two-person security team used AI to find real vulnerabilities in code and earned $250,000 from a single bounty. Now they have open-sourced the system behind it. It is called Open-Kritt. Most people scan a repository by giving an AI one giant task: “Find every vulnerability in this code.” Open-Kritt takes a more focused approach. First, one agent maps every reachable entry point. Then it launches separate agents to trace each code path and search for specific failures. Each agent gets its own isolated container, where it can: > Inspect the code > Install security tools > Compile the project > Run tests > Fuzz inputs > Build a proof of concept The raw findings are then checked, merged when they describe the same bug, and ranked so researchers can inspect the most promising ones first. It works with Codex, Claude Code, OpenAI, Anthropic, and OpenRouter. The Kritt team says it used the system to find more than 20 High and Critical vulnerabilities in nine months. It also finished first in the Firedancer V1 audit competition using fully AI-driven research with no manual review. This is an open-source research lab that turns AI agents into a small security team. github.com/Kritt-ai/open-…

@Prithvir12 This feels right. General humanoids are the demo; bounded task cells are the business.

humanoids today feel like L1s in 2020 fat valuations abundant capital real technical progress but de minimis revenue humanoids are not commercially useful. we need: + fixed morphology + bounded environments + deterministic tasks + closed-loop perception and control and the pitch can't be “cheaper humans.” it needs to be machines that operate beyond human limits: + continuous uptime + sub-millimeter precision + millisecond reaction + mass coordination + hostile environments humanoids are L1s specialized machines are stablecoins. generality captures imagination constraint creates value.

@gdb I’ve been using Remote a lot

ask chatgpt work to do any recurring task

@nitishkmrk Stress-testing visual consistency at 10k samples is the right instinct. Coding agents need more of this: not just ‘it worked once,’ but repeatability under ugly volume

i built 24 Codex icon skills and stress tested them across 10,000 icons for visual consistency, here is quick timelapse ⎯⟡°

@sytaylor The interesting competitor to Unreal may not start as an engine. It might start as a world-spec compiler: intent in, playable constraints out, then only later become a full asset/runtime stack

What's impressive about this is not the game quality It's the fact it builds this whole, working world with no assets or MCP to pull from. If that's what the base models do... Who's building the competitor to unreal engine that works this way?

@ciprianveg How much was this setup?

Preparing 16xGB10 cluster for Kimi K3, Deepseek V4 Pro, and future GLM 5.5 and Minimax M4..

@yasser_elsaid_ The folder-per-system pattern is underrated. Makes the agent treat each business surface like an inspectable codebase instead of a magic integration. I’d still want a thin audit log around every write action though

If you're a startup founder, please connect google ads, meta ads, smartlead, apollo, chatbase (if you use it), attio (or your crm) to claude/codex through cli/api (not mcp). Make each of those tools a folder inside one main repo. Start a session inside each one or in the parent repo if you want them to talk to each other. Have your agent write code so that when you ask for a "pulse" it fetches all data from all the platforms and tells you what's good, what's not and what you need to fix. BUT make sure to fully understand the data and don't say yes to every recommendation from LLMs, LLMs are still not good at considering other variables they dont see, not good at correlation/causation, and jump to conclusions too quickly. This setup is amazing for your to actually understand your business, what your customers want and what the market responds to and be able to run a big part of your gtm conversationally.

@rohanpaul_ai Judgment is the part people keep trying to benchmark with task completion

AI agents given 6 days and $3K produced two research papers, and both were rejected. The people who had spent months on those questions graded what the AI agent wrote. The failure was judgment. The main runs used Claude Opus 4.8 with extra-high reasoning on the OpenClaw scaffold, chosen after dry runs across OpenAI and Anthropic models, including an early pilot with GPT-5.3 Codex that could not handle the scaffold. Execution was never the problem. The agents ran hundreds of experiments, debugged crashing GPU pods, and compiled camera-ready LaTeX without a human touching anything. They were honest about it too, because the logs show marketable claims being retired in favor of negative results rather than any reward hacking. The failure was judgment. Round after round of automated reviews came back negative, but each response narrowed the claim and added a caveat instead of redesigning the experiment. Neither run noticed it was short on ideas rather than money, since both ended with over half of the $3K unspent. – arxiv. org/abs/2607.27191 Title: "Can AI agents conduct open-ended AI research? Early evidence from two case studies"

@the_osps everse engineering is a strong MCP use case

Ghidra MCP Server provides 256 MCP tools for AI-powered reverse engineering with Ghidra integration.

@worktreewise This is one of those boring dev-ops details that suddenly matters a lot with coding agents

Stop re-running local database migrations every time you change branches. Use worktrees with isolated local DB instances for total development freedom. #Git #GitHub #DevTools

@PedroLaRosaDev Agents reviewing agents is useful if the artifact is structured: risk summary, files touched, assumptions, tests run, tests missing

Claude Code just made /code-review a background subagent in V1.2.218 Agents reviewing agents while you keep working. We're one release away from them leaving passive-aggressive PR comments on each other. 😏 Source: claude.ai/chat/22fa8437-…

@stretchcloud Stateless MCP feels like the right direction for internet-native agents

The biggest change to MCP since launch just dropped, and I think the protocol layer conversation is about to get a lot more interesting. MCP 2026-07-28 goes stateless. The initialize handshake is gone. No more session IDs. Every request is now self-contained: it carries protocol version, client identity, and capabilities in the request itself. What this unlocks: any request can hit any server instance behind a plain round-robin load balancer. You can now run MCP servers as stateless Lambdas or Cloud Run containers. That removes the main scaling constraint that made hosted MCP servers complicated to deploy. The spec also introduces Tasks: a formal model for long-running work, where a server returns a task handle and the client drives completion via tasks/get, tasks/update, tasks/cancel. Big deal for coding agents doing multi-step work. SDK releases: Python v2 (FastMCP renamed to MCPServer), TypeScript v2, Go v1.7.0-pre.1, C# 2.0.0-preview.1. AWS AgentCore Gateway already supports the spec. Simon Willison's immediate reaction: shipped mcp-explorer and datasette-mcp. His prior reluctance was context pollution. Stateless removes that. My take: MCP was always the right abstraction. Stateless is what makes it production-ready. x.com/simonw/status/…

@ctushr Agree. The review surface changed once agents started producing bulk diffs. The missing tool is less ‘AI reviewer’ and more an operator console: intent, risk map, test evidence, ownership, and what changed since the last known-good state

Baffling how many AI code review tools exist now and not one solid human-focused reviewing tool. Reviewing stopped being a second-class citizen the moment agents started writing the code. Honestly, the tooling hasn't caught up.

@heyyyyyieeee Let’s connect!

If you're in tech 🧑🏻‍💻 Let's connect and grow together 🤝✨

@thsottiaux So basically weekends are for casually changing science.

The week was for efficiency. The weekend is for 10 major breakthroughs in science. There will be signs.

@daradoescode @vercel Agent-first design is underrated here

I've been working on Weaver an OSS mac/windows desktop widget framework built on @vercel's Native platform Here's what makes it cool - .tsx w/ tailwind like syntax for authoring with out electron or webview - Agent first design 'weaver check' to statically find errors n hot reload - Widget logic is crash isolated - Declarative permission system There's still a lot left to work on: more perf improvements(mac os isnt where i want it to be), more primitives for building better widgets, shareable widgets w/ agentic remixes and a manager I've been working on it for about 2 weeks n its crazy to think this would have been impossible 3 years ago. I spent 7 billion tokens over the last month it wouldn't have been possible w/ out OSS programs from @OpenAI n @greptile OFC i won't forget resets from @thsottiaux 🤑🤑

@gdb Same here. It’s been really good for going deep without losing track of the details.

having a great experience with using chatgpt work for extremely thorough research and fact checking

@daniel_mac8 The default knob problem is real. A lot of ‘model quality’ complaints are actually orchestration/config issues: reasoning effort, context shape, test loop, and when the agent is allowed to stop

Codex users: do this right *now*. Max reasoning effort is off by default. Luna at Max reasoning is ~ Sol Medium / Opus 5 Medium level at 1/6th the cost. Change your life today.

@ASvanevik @sama Codex feels strongest when it stays close to the repo and keeps momentum.

codex is remarkably good. @sama should triple-down on it.

@Krishnasagrawal Free source material from the builders beats another repackaged course every time.

🚨Anthropic just showed a 24-minute workshop on how to actually do prompts for Claude. Taught by the people who built it. Free. No registration. No paywall. I've seen $300 courses that don't cover what they teach in the first 8 minutes. Watch it and bookmark it now! Video Credit : Respective owner