Bilingual

Monetization Gateway: Charge for any resource behind Cloudflare via x402
20 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
19 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
19 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
19 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
22 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
19 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
19 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
18 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
18 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
18 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.
Context graphs: how AI agents can store and use past decisions
16 days ago
- Context graphs store the 'why' behind decisions in an agent's memory, capturing reasoning and connections between entities, which traditional systems fail to record.
- Flat context windows cause context rot and lack decision traces, leading to inefficiencies as LLMs struggle with large, unstructured data and miss critical tribal knowledge and past precedents.
- Context graphs structure memory as nodes (entities) and edges (relationships), enabling agents to traverse structured data efficiently, reduce token usage and latency, and avoid recomputing links.
- Decision traces within context graphs record problems, options, constraints, exceptions, reasoning, and outcomes, turning past decisions into precedents that agents can use for autonomous learning and improvement.
- Implementation involves capturing decisions at the moment they are made to minimize friction, using graphs to cache hops and enforce reasoning storage, and integrating with an orchestration layer for full context across systems.
- Agentic search, while effective for correctness, incurs high costs and latency due to repeated LLM calls and cannot retrieve unreasoned data, whereas context graphs optimize these factors and mandate reasoning capture.
- Challenges include ensuring high-quality rationale input, avoiding decision swamps through curation, and balancing human vs. agent inference, as the technology is still early but shows promising results.
- A full AI-native stack with context graphs includes systems of record, a harness for execution, the graph for decision traces, and agents/humans in a feedback loop, enabling universal context and continuous learning loops.
- Context graphs are particularly valuable for automating complex, exception-heavy, cross-functional processes like procurement, claims, and compliance, where decisions depend on nuanced, multi-system context.
HTML as a native data format for LLMs
16 days ago
- Using HTML instead of JSON for document templates made AI agents more efficient and reliable by leveraging the model's pre-trained fluency in HTML.
- A single tool for whole-tree rewrites (set_template_markup) reduced round trips, prevented intermediate invalid states, and ensured coherent edits.
- HTML's external labeling (e.g., data attributes) makes the tree readable for both humans and AI, unlike JSON's nested, unlabeled structure.
- A surgical tool (set_node_attributes) allowed minor tweaks without full rewrites, utilizing browser-native APIs for simplicity.
- Key lessons include preserving identifiers rigorously, validating at boundaries, and storing typed data internally while exposing HTML only at authoring edges.
- Reference implementation barkup enforces round-trip integrity, declared type coercion, and structured error handling to avoid common pitfalls.
GitHub Under Siege
15 days ago
- GitHub's reliability has declined with frequent outages, especially in Actions, disrupting many teams' workflows.
- There is a growing trend of developers and companies leaving GitHub due to platform lock-in, governance concerns, and centralization risks.
- Competitors like Cursor's Origin, Entire, and autonomous-agent platforms are emerging, challenging GitHub's dominance by focusing on agent-native workflows.
Karpathy, Google, Tan agree Markdown is the answer, but not for the same problem
14 days ago
- Andrej Karpathy's 'LLM Wiki' proposes using linked Markdown files for AI agent memory.
- Google introduced Open Knowledge Format (OKF v0.1), packaging organizational knowledge in Markdown for enterprise use.
- Garry Tan's gstack uses Markdown files to define specialist coding agent roles, achieving high GitHub popularity.
- Markdown has become the standard format for agents to read and write, evolving from developer conventions like CLAUDE.md.
- Competitive advantage shifts from model ownership to owning accumulated Markdown knowledge bases that are portable across systems.
- The durability of Markdown standards like OKF depends on adoption and ecosystem development to avoid remaining just ideas.
Apple Silicon Exec Explains Mac Mini AI Demand and On-Device Future
15 days ago
- Apple's Mac mini and Mac Studio are favored for running AI agents due to demand for controlled, isolated systems that run continuously.
- AI tools being Mac-first or Mac-only strengthens Mac's position among developers, including those at frontier AI labs.
- Agentic AI is viewed as a whole-chip problem, leveraging Apple silicon's strengths beyond just GPU for tasks like tool-calling.
- Apple's AI strength traces back to early chip decisions, including the Neural Engine and neural accelerators in CPU and GPU.
- A shift toward local AI is driven by privacy, security, and cost concerns, with a hybrid future envisioned for on-device and cloud tasks.
- Transparent AI on iPhone and iPad includes features like Draw Things and SwingVision that operate seamlessly without overt AI labeling.
- AI development is progressing rapidly, with unpredictable advancements expected in the near future.
How Version Control Will Evolve for the Agent Boom
12 days ago
- Git will remain popular but must evolve for AI agent-driven development.
- Storing agent session logs with code captures context and intent, improving accuracy and review.
- Decentralized Git hosting is needed to avoid bottlenecks from centralization and support agent scalability.
- Developers will orchestrate agents, using Git as a source of truth for coordination and collaboration.
Introducing Muse Spark 1.1
12 days ago
- Meta launches Muse Spark 1.1, an upgraded multimodal reasoning model for agentic tasks with improved tool/computer use, coding, and multimodal understanding.
- The model features a 1-million token context window, multi-agent orchestration for faster project handling, and public preview via the Meta Model API.
- Muse Spark 1.1 excels in computer-use workflows, coding (debugging, migrations), multimodal perception, and safety, with strong adversarial robustness.
- Early partners like Replit and Box praise its comprehensive capabilities for large-scale agentic workloads and enterprise applications.
- Meta highlights ongoing research momentum, with more advanced models in training, advancing toward personal superintelligence.
Why we're moving off Cloudflare Durable Objects
12 days ago
- Wire, a context container service for AI agents, is moving off Cloudflare Durable Objects due to four structural limits.
- Limitations include the vector index being separate from the object, compute not being next to data, placement being static at creation, and inability to self-host.
- Wire built a new runtime using Bun on Fly Machines, with each container as an SQLite file embedding a vector index via sqlite-vec for in-process retrieval.
- The new system offers improved retrieval performance, placement flexibility, and better isolation for paid workspaces.
- Durability is ensured through continuous WAL shipping, and the transition maintains the same API and user interface.
Build or buy an agent developer workspace?
12 days ago
- Monaco initially experimented with a 'Million Interns' initiative using tmux and MCP servers to manage multiple AI agents but faced issues like developer interface preferences, port conflicts, and resource contention.
- After exploring external agent orchestration platforms like Ramp's Inspect and Stripe's Minions, Monaco found gaps such as high costs, lack of a populated data plane, and challenges in consolidating local service setups.
- Monaco adopted Coder to create 'Monacoder,' an in-house solution featuring cloud-hosted workspaces with 1:1 app-to-VM cardinality, baked-in dependencies via AMIs, and secure Claude access through Secrets Manager.
- Key technical decisions included using Docker Compose for runtime, seeding the data plane with de-identified sales data, managing databases with Postgres Kubernetes Operator, and integrating MCPs like Linear and Datadog.
- Monacoder has been widely adopted, with many developers using 3-6 workspaces simultaneously, and it supports workflows like automated PR review and CI failure addressing, accelerating code quality and security improvements.
- The system has proven low-maintenance with minimal costs, and extensions have been added for Linear delegation and GitHub webhooks, with plans to further integrate with Slack and other developer tools.
- Despite competition from tools like Cursor and Niteshift, Monaco sees value in building in-house, vendor-agnostic solutions to keep pace with evolving AI technologies, and is hiring for developer experience roles.
I think I was part of a model distillation attack
11 days ago
- The author received emails from OpenAI about inconsistent API usage and a 'Prohibited Biological Use' policy violation.
- Upon investigation, they found 114 unauthorized interactions with gpt-5.5 using a compromised legacy API key from 2023, possibly from bettergpt.chat.
- The attackers used the key for various tasks including a video QA benchmark, coding agents, and exporting insulin molecule chains, which triggered the policy violation.
- The attack involved a repo named 'ga-synthesis' where agents ran in environments like PyMOL, interacting with GUI apps via virtual desktop.
- The system prompts indicated usage of Anthropic's Claude Agent SDK, with paths suggesting a Chinese lab on Alibaba Cloud.
- The attack lasted about 10 minutes, used 4M tokens, cost $23, and was stopped by disabling the key after noticing it at midnight.

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

Context graphs: how AI agents can store and use past decisions

HTML as a native data format for LLMs

GitHub Under Siege

Karpathy, Google, Tan agree Markdown is the answer, but not for the same problem

Apple Silicon Exec Explains Mac Mini AI Demand and On-Device Future

How Version Control Will Evolve for the Agent Boom

Introducing Muse Spark 1.1

Why we're moving off Cloudflare Durable Objects

Build or buy an agent developer workspace?

I think I was part of a model distillation attack