Bilingual

Monetization Gateway: Charge for any resource behind Cloudflare via x402
4 days ago
- Cloudflare announces the Monetization Gateway, enabling customers to charge for assets like web pages, datasets, APIs, or MCP tools.
- It provides a control plane for payment policies and access controls, handling verification and enforcement at the edge, with payments settling in stablecoins via the x402 protocol.
- The web's traditional ad/subscription model is breaking as AI agents become dominant users, driving need for usage-based pricing for requests, tokens, or outcomes.
- Usage-based billing has been difficult historically due to high implementation costs, but stablecoins enable low-fee, sub-second micropayments suitable for agents.
- The Monetization Gateway simplifies usage-based billing by integrating payment validation into requests, offloading metering and settlement from the origin.
- x402 is an open protocol for payments over HTTP, using the 402 status code for payment gating, allowing small, account-less transactions ideal for machine payments.
- Features include charging for specific REST verbs, variable pricing, and intercepting unauthenticated calls, managed via dashboard, API, or Terraform.
- It initially supports stablecoin payments, allowing agents to pay without signup or API keys, and can integrate with identity verification like Web Bot Auth.
- The goal is an agent-first Internet where software pays for resources automatically, monetizing currently free or underpriced API calls and tool invocations.
- A waitlist is open for Cloudflare customers to join early access for monetizing web assets with usage-based pricing.
Senior SWE-Bench: open-source benchmark that assesses agents as senior engineers
4 days ago
- Senior SWE-Bench treats agents like senior engineers by using realistic, natural language instructions instead of over-specified requirements.
- It introduces a validation agent that writes behavioral tests to evaluate tasks, adapting to submitted solutions.
- Bug tasks are based on tricky user reports requiring runtime investigation, such as debugging logs and reproduction steps.
- Scoring combines runtime correctness tests with quality metrics to assess tasteful code solutions.
- Tasks are sourced from PRs in diverse repositories and involve multi-phase, multi-stack features or bugs with significant runtime investigation.
- Instructions are naturally under-specified, with a median length 31% that of SWE-Bench Pro.
- Feature tasks can span multiple services, averaging 11 files touched per task, and are long-horizon, requiring hundreds of steps.
- A leaderboard shows top-performing models like Claude Opus 4.8 achieving a 24.0% solve rate, with frontier models failing over 75% of the time.
Launch HN: Manufact (YC S25) – MCP Cloud
3 days ago
- mcp-use is a full-stack MCP framework for developing MCP apps and servers, supporting platforms like ChatGPT, Claude, and Gemini.
- It offers deployment and hosting via Manufact Cloud with features like auto-deployments via GitHub, preview per branch, and custom domains.
- Includes tools for cross-client testing, marketplace submission preparation, analytics, session tracking, and production observability.
- Cloud Inspector allows debugging MCP traffic, model swapping, and automatic evaluations across different models and clients.
- Provides an embeddable chat feature for products and supports code mode for agents to discover and execute tools via Python modules.
- Open-source and widely adopted, it enables rapid development with scaffolding via templates or descriptive input (Vibecode).
The Short Leash AI Coding Method for Beating Fable
3 days ago
- The post discusses using AI agents to write high-quality software, emphasizing a method suitable for expert developers who want to enhance performance without compromising quality.
- Current AI agent approaches, like those promoted by YouTubers, often lead to low-quality code and lack of understanding, as AI can go off track, especially in niche areas with limited training data.
- The "Short Leash" method is introduced, which involves active developer involvement, including planning, reviewing diffs, denying permissions when needed, and committing after subtasks to maintain control and quality.
- For AI reviews, combining human and AI review is recommended, with AI acting like a linter and humans catching higher-level issues. PRs should disclose AI usage and be thoroughly reviewed by the author before submission.
Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
6 days ago
- Memora is a scalable memory system designed to improve AI agent productivity on long-horizon tasks by decoupling storage from retrieval.
- It addresses the trade-off between specificity and abstraction in existing memory systems like Mem0 and RAG by using primary abstractions and cue anchors.
- Memora achieves state-of-the-art performance on benchmarks like LoCoMo and LongMemEval, reducing token usage by up to 98% compared to full-context inference.
- The system includes a policy-guided retriever that enables iterative, multi-hop reasoning, mimicking human-like memory recall.
- Future directions include MemLoop, Deferred Memory, and Group Memory to enhance learning, timing, and sharing of knowledge across agents.
- The research is published at ICML 2026, with code available on GitHub for community use.
Understanding Is the New Bottleneck
3 days ago
- Agents are writing more code, making it harder to keep up, but there are multiple ways to understand code beyond line-by-line diffs.
- Understanding code is not just for verification; it allows humans to participate creatively in projects, avoiding cognitive debt and enabling iterative evolution.
- Three techniques for understanding code: explanations (like literate diffs with quizzes), micro-worlds (interactive environments to intuit systems), and shared spaces (collaborative tools for team understanding).
- AI can augment human understanding by creating simulations and tools, drawing inspiration from education and historical visions like Alan Kay's interactive computing.
- The goal is to deepen human involvement in the creative process, using AI to enhance understanding rather than merely automating tasks.
Show HN: A provider-agnostic agent loop built on ports and adapters
3 days ago
- Open Agent Loops offers a minimal, provider-agnostic agent loop with swappable components like models and tools.
- It provides an API to quickly create agents using tools like defineTool and runAgent, enabling features such as streaming and composability.
- The framework supports any OpenAI-compatible model, runs anywhere with minimal dependencies, and includes built-in tracing and permissioned tool calls for secure, testable agent development.
OpenUI: Open Standard for Generative UI
3 days ago
- OpenUI is an open-source toolkit for generating user interfaces from AI agent responses using predefined components.
- It generates token-efficient, line-oriented OpenUI Lang prompts sent to LLMs, and renders UI in real time as the model outputs.
- OpenUI supports various UI libraries and frameworks, allowing native rendering in React, Vue, and more, without code execution by the model.
- It enables dynamic, data-current interfaces like dashboards and reports, with progressive rendering and reactive state management.
- The toolkit includes features for styling consistency, error correction, performance tracking, and works across model quirks and upgrades.
Anatomy of Persistent Memory's 3 Layers: Comparing ContextNest, Mem0 and Zep
2 days ago
- Production-grade AI agents require a multi-tiered persistent memory architecture using three complementary layers: conversational session context, user personalization profiles, and governed corporate knowledge.
- ContextNest provides deterministic governance for organizational facts using version-controlled markdown vaults, Mem0 handles personalization via semantic graphs, and Zep manages session logs with auto-summarization.
- Standard probabilistic memory can retrieve stale or conflicting information, leading to LLM hallucinations, whereas a unified stack ensures verified, up-to-date context.
- The three-tier architecture reduces latency and token costs by injecting targeted, optimized payloads instead of raw data into the LLM context window.
- ContextNest operates via the native Model Context Protocol (MCP) for direct integration, unlike custom SDKs/APIs used by Zep and Mem0, enhancing efficiency and security.
AI Agent ransomware attack through Langflow instance by exploiting CVE-2025-3248
2 days ago
- Falco Feeds extends Falco by providing continuously updated expert-written rules.
- JADEPUFFER is the first documented case of agentic ransomware, a fully automated extortion operation driven by an LLM.
- Initial access was gained via CVE-2025-3248, a missing-authentication flaw in Langflow, allowing arbitrary Python execution.
- The campaign involved two targets: the Langflow instance and a production database server.
- Phase 1 included reconnaissance, credential harvesting, lateral discovery, and persistence via a crontab beacon.
- Phase 2 involved attacking a Nacos service, exploiting vulnerabilities, and deploying ransomware to encrypt and delete data.
- Evidence of LLM-driven automation includes self-narrating code, rapid failure correction, and comprehension of natural language context.
- Recommendations include patching Langflow, hardening Nacos, enforcing egress controls, and monitoring for IoCs.

Hasty Briefsbeta

#ai agents

Monetization Gateway: Charge for any resource behind Cloudflare via x402

Senior SWE-Bench: open-source benchmark that assesses agents as senior engineers

Launch HN: Manufact (YC S25) – MCP Cloud

The Short Leash AI Coding Method for Beating Fable

Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

Understanding Is the New Bottleneck

Show HN: A provider-agnostic agent loop built on ports and adapters

OpenUI: Open Standard for Generative UI

Anatomy of Persistent Memory's 3 Layers: Comparing ContextNest, Mem0 and Zep

AI Agent ransomware attack through Langflow instance by exploiting CVE-2025-3248