440 things worth a look, from across the AI agent world.
Helps developers and AI agents maintain interconnected markdown docs that sync with code via wiki links and backreferences.
Orchestration platform running Claude Code and Codex locally or in cloud, with git worktree isolation and visual diffs.
No-code automation platform with 280+ integrations, an extensible TypeScript pieces framework, and MCP server support.
A Claude Code plugin bundling specialized agents, skills, hooks, and rules to optimize agentic workflows across multiple AI dev harnesses.
An open, lightweight event-based protocol standardizing how AI agents connect to user-facing applications in real time.
Production infrastructure to build, run, and scale AI agents as callable APIs with observability, governance, and audit trails.
A Python agent using Claude's API with file and bash tools to autonomously complete tasks through an interactive CLI.
Bootstrap and run a multi-agent AI development team for planning, development, testing, and deployment across parallel chats.
A TypeScript course teaching production AI apps with Vercel's AI SDK, covering streaming, tool calling, and multi-provider support.
Workshop with practical exercises on LLM evaluation, including LLM-as-a-judge scoring of AI-generated responses.
Syncs data from 50+ integrations and exposes it through a unified, LLM-friendly search interface for agents and RAG.
A skill using mitmproxy to capture and inspect Claude Code's HTTPS traffic for debugging, analysis, and optimization.
Converts PDFs and image documents into clean text and markdown using a 7B vision language model, handling tables, equations, and handwriting.
A behavioral framework helping Claude avoid common coding errors via clear thinking, simplicity, and surgical edits.
Enforces filesystem and network restrictions on processes via OS primitives (sandbox-exec, bubblewrap) without containers.
A Jupyter notebook from Anthropic's cookbooks demonstrating how to use Claude for text classification tasks.
A guide to using Claude's Skills feature for document generation, data analysis, and business automation with examples.
An interactive notebook tutorial on prompt engineering for Claude, from basic structure to few-shot prompting and tool use.
Reference implementation for finding and fixing code vulnerabilities using Claude agents for threat modeling, scanning, and patching.
MCP server that streams interactive Excalidraw diagrams with viewport control and fullscreen editing for Claude and other clients.
An automated feedback loop that tests voice AI agents against adversarial scenarios and keeps prompt changes that improve metrics.
A curated collection of books, courses, papers, tools, and datasets for studying AI and machine learning.
An AI agent system enforcing approval gates at each step, combining planning, approval, and execution with human control.
Autonomous experiment framework that improves metrics by trying ideas, keeping wins, reverting failures, and logging experiments.
Open-source framework for orchestrating multiple AI agents with intelligent routing and context management in Python and TypeScript.
AI skill that integrates security best practices into Claude and other coding assistants to prevent shipping vulnerabilities.
A Claude Code stop hook that keeps the agent working until all plans and user requests are fully complete, requiring an explicit done token.
A dockerized OCR solution converting PDFs to Markdown using DeepSeek-OCR, with batch scripts and REST API endpoints.
Fullstack template combining a Mastra backend with a React frontend and OpenAI-powered AI capabilities.
A from-scratch 1.5B GPT with FineWeb pretraining, horror fine-tuning, and GRPO reinforcement learning for scary stories.
Exposes Claude conversation history as a Durable Stream with a real-time web UI for viewing sessions live.
Connects LLMs to Chrome via CDP to automate browser tasks with a self-improving harness agents can edit at runtime.
Rust-based terminal interface combining LLM control, Chrome automation, and a steerable UI for browser task automation.
A customizable AI-powered command-line coding tool leveraging Groq's API for fast iteration.
Framework for building SaaS apps where AI agents and user interfaces operate as equal partners in shared, real-time environments.
A TypeScript self-hosted app converting files between 1000+ formats using FFmpeg, ImageMagick, Pandoc, and LibreOffice.
A reinforcement learning framework for efficient LLM post-training with distributed parallelism and multi-turn rollouts.
Resources and materials for Chip Huyen's 2025 book 'AI Engineering', covering foundation models, LLMs, and AI applications.
Compresses tool outputs, logs, and RAG chunks 60-95% before reaching LLMs, as a library, proxy, or MCP server.
A DSPy module demo using Chroma Cloud for semantic search to augment an LLM's responses.
MCP server enabling AI-driven parametric design automation in Onshape via its REST API.
Dispatches LLM tool calls mid-stream rather than after completion, cutting agent latency 1.2-1.5x via overlapped execution.
Analyzes codebases to generate humorous yet detailed COBOL migration assessments with timelines, costs, teams, and risk analysis.
Open-source AI tool using specialized agents to understand projects and make precise code edits through terminal commands.
An open-source testing framework for LLM apps with metrics for relevancy, hallucination, and task completion.
Demonstrates integrating CopilotKit's canvas component with the Mastra AI framework for interactive applications.
A Next.js demo app showcasing MCP server-client integration for managing projects and tasks with AI assistance.
A TypeScript utility that builds structured LLM prompts by combining context fields like task, tone, examples, and formatting.
RL infrastructure for training LLM agents on long-horizon terminal and coding tasks, topping the TerminalBench leaderboard.
An AI coding agent extension running autonomous optimization loops to improve metrics like speed, bundle size, or loss.
Creates compiled functions for high-performance redaction of sensitive object properties with minimal stringify overhead.
A collection of ready-to-use configurations, agents, commands, and MCPs to enhance workflows with Claude Code.
Agent skill guiding iterative creation of three-tier docs with author reviews at each step.
Blocks dangerous git and shell commands from running in AI coding agents like Claude, Codex, and Copilot before they destroy work.
AI agent plugin that minimizes code generation by reusing existing solutions, cutting output 80-94%.
A bash script wrapping macOS caffeinate to keep a Mac awake for a duration, while apps run, or until a set time.
Global pi config workspace with npm-managed extensions for auto-discovery and dependency management.
Fast, accurate file search library for AI agents, Neovim, Rust, C, and Node.js with frecency ranking and typo resistance.
Orchestrates multiple coding agents in parallel with isolated git workspaces for reviewing and deploying AI code.
A Python system deploying multiple specialized AI agents in parallel for comprehensive, multi-perspective analysis.
An autonomous system that optimizes code and ML models by iteratively testing ideas, measuring metrics, and keeping wins.
Fork of nanochat that runs on Modal serverless, training a ChatGPT-like model on 8xH100 GPUs for under $100.
Open-source implementation of Vercel's v0 agent that helps developers design, build, and ship applications using AI.
Framework that evaluates multiple document-splitting methods via quality metrics to select the optimal chunking approach per document.
An AI-powered tool that analyzes and refactors Python code while preserving behavior, using DSPy and deterministic checks.
Lightweight agent framework using DSPy modules with CLI access, a FastAPI server, and eval tools for OpenAI and Ollama.
Plugin of 37 skills and 51 agents structuring AI-assisted development around planning, execution, review, and knowledge docs.
Free open-source collaborative whiteboard for hand-drawn style diagrams with end-to-end encryption and real-time collaboration.
Open-source ESLint plugin with custom lint rules designed to guide AI coding agents toward better output quality.
Hardware implementation of a small GPT in Verilog on a Xilinx Virtex-5 FPGA generating ~56k tokens/sec.
A drag-and-drop interface for building AI agent pipelines with real-time execution, web scraping, and human-in-the-loop approval.
A framework for creating mobile, desktop, and web apps using only Python, with no frontend experience required.
An open-source visual programming tool integrating AI workflows and backend logic into TypeScript codebases via VSCode.
Research repo formalizing context engineering as entropy reduction, tracing its evolution from GUIs to superhuman AI systems.
An MIT-licensed toolkit of 23 specialized AI agents and power tools automating the full software development lifecycle.
Transforms MCP servers into Claude Skills with progressive disclosure, dramatically reducing token consumption.
Fast offline LLM inference engine in ~1,200 lines of readable Python with prefix caching and tensor parallelism.
An async Python agent integrating Gemini with Pipedream's MCP server for tool-assisted Gmail and Google Calendar tasks.
A skill for building real-time bidirectional streaming apps over WebSockets with Gemini for audio, video, and text.
A desktop app for running multiple AI coding agents in parallel with isolated Git worktrees across providers.
A framework for optimizing prompts, code, and agent architectures via LLM reflection and Pareto-efficient evolutionary search.
A framework using LLM reflection and Pareto-efficient evolutionary search to optimize prompts, code, and agent designs.
An open-source framework that builds and queries temporal knowledge graphs tracking how facts evolve over time.
Workshop teaching how to build an AI coding agent using the Claude API and Go, from basic chat to file editing and code search.
An LLM-agnostic memory layer that turns agent execution and conversations into structured, persistent state for production.
Examples for building voice agents with the Gemini Live API, processing audio, video, and text in real time.
Multi-tenant vector and full-text search engine using tiered RAM/NVMe/object storage for low-latency queries on cheap storage.
Provider-agnostic, open-source evaluation infrastructure offering 30+ standardized benchmarks for any model provider.
A Chrome extension that exposes browser functionality via the Model Context Protocol for AI-controlled automation.
Generates animated educational videos from text prompts using Manim, visualizing math concepts, algorithms, and data.
Fully-automated framework letting users create and deploy LLM agents through natural language alone, with no coding.
An MCP server and skill that lets LLMs reason iteratively over large codebases by keeping working state in a Python process outside the prompt.
An RL agent that trades Polymarket binary crypto prediction markets using real-time Binance futures and orderbook data, training on-device with MLX.
An SDK implementing the MCP Apps standard to build interactive web interfaces for AI tools in TypeScript, Ruby, and Python.
A Jupyter notebook series teaching deep learning fundamentals through practical projects, from neural networks to GPT models.
An open-source kernel providing durable, inspectable work records for multi-agent collaboration across humans and tools.
Open-source desktop app and TypeScript library for building complex AI agents, prompt chaining, and LLM integrations.
Async system using LLMs to generate, test, and refine instructions for solving Abstraction and Reasoning Corpus challenges.
A multi-agent, multi-channel orchestrator for the one-person company that runs isolated AI agent teams collaboratively, formerly named TinyClaw.
Lightweight native Linux sandbox isolating agent code execution with filesystem, network, and resource limits, no containers.
An open-source multi-agent framework handling complex data science tasks via iterative planning, execution, validation, and refinement.
An AI tool that generates publication-ready scientific documents with real-time literature research and verified citations.
An AI agent built with the Claude Agent SDK that trains state-of-the-art ML models using scientific agent skills.
A categorized collection of over 120 libraries covering LLM training, inference, RAG, agents, and evaluation.
A minimal, dependency-free GPT implementation in ~200 lines of pure Python covering tokenization, autograd, attention, and training.
A minimal LLM training framework for training GPT-2-capability models on single GPU nodes for under $100.
TypeScript Workflow DevKit for AWS enabling serverless, resumable workflows using Lambda, DynamoDB, SQS, and S3.
An AI system that builds and improves agent implementations overnight by modifying prompts, tools, and config while benchmarking.
An MCP server letting AI clients execute tasks across email, calendars, and task managers while keeping keys and data private.
Autonomous AI pentester that analyzes source code, finds vulnerabilities, and executes real exploits to prove security flaws.
A drag-and-drop interface for building AI agents, workflows, and tools without coding, generating production TypeScript.
An AI agent that autonomously builds verified knowledge graphs for personalized sales outreach with RL feedback loops.
A Git-style conversation forking chat app that lets you branch conversations at any point to explore different paths with AI.
A batteries-included agent harness with filesystem access, sub-agents, and context management for any tool-calling LLM.
An asynchronous coding agent platform for building custom AI software engineering agents, integrated with Slack, Linear, and GitHub.
Pre-built evaluation tools for assessing and measuring the performance of LLM applications.
A graphical interface for creating and running LangGraph workflows with local or cloud-based LLMs.
A CLI tool and standards framework for building production-ready AI agents with project templates, testing, and prompt versioning.
Lets you push Git repos to your own storage backend like S3, R2, Tigris, MinIO, Postgres, SFTP, or local disk.
Build AI agents with persistent advanced memory that can learn and self-improve over time, via CLI and API.
A Next.js and React voice AI frontend for real-time interaction with LiveKit Agents, with video, screen sharing, and audio visualizers.
An observability platform built in Rust providing tracing, evaluations, monitoring, and SQL access for AI agent apps.
A local app extending Codex with persistent loop modes, completion checks, and Telegram/Slack notifications for long tasks.
A personal website and blog built as a server-side rendered Rust app on Axum with hot-reloading content.
Native Swift/AppKit terminal with vertical tabs, notifications, a built-in browser, and CLI scripting for AI agent workflows.
Native iOS app that syncs personal tasks with GitHub Issues through a SwiftUI interface authenticated via GitHub OAuth.
Demo project showcasing Mastra with AI UI frameworks (AI SDK, Assistant UI, CopilotKit) and patterns like generative UIs.
Claude skill that teaches a user a new concept over multiple sessions, using the current directory as a stateful teaching workspace.
MCP server letting AI agents use the full Cloudflare API by executing code against the spec, avoiding huge context overhead.
A self-hostable personal finance app built with Ruby on Rails and PostgreSQL, now archived under AGPLv3.
An MCP extension spec enabling servers to deliver interactive HTML UIs to hosts via a ui:// URI scheme with bidirectional JSON-RPC.
A curated collection of papers, frameworks, and guides on the evolution from prompt engineering to production AI systems.
A Gymnasium-style framework for creating and using isolated execution environments to train AI agents via reinforcement learning.
Easy-to-use BI and embedded analytics tool letting teams ask questions and learn from data without SQL knowledge.
Connects AI models to thousands of APIs and tools via MCP through single function calls for enterprise-scale deployments.
Pi extension where the model writes JS to fan out work across isolated subagents, then synthesizes results.
Autonomous bash script that runs various AI agents in a loop to complete tasks from a PRD until finished.
A framework for optimizing AI agents with reinforcement learning, compatible with any agent framework and requiring minimal code changes.
Open-source solution for automated AI-agent phone calls with multi-language support and Azure integration for call centers.
An AI-powered tool to explore data and generate interactive visualizations through natural language queries.
Framework for building AI-powered agents and extensions for Microsoft Teams with a simplified architecture and better DX.
Build AI workflows with specialized agents, automatic handoffs, and coordination for the AI SDK v5.
Essential utilities for building production AI apps with the Vercel AI SDK, including state, debugging, and agents.
Converts command-line tools into typed JavaScript APIs by parsing help text, letting agents call CLIs with structured interfaces.
A tool that validates agent-generated code in real browsers via Playwright, detecting performance, security, and design issues.
An open-source AI platform enabling knowledge workers to automate tasks, build internal tools, and deploy AI anywhere.
An 8B-parameter text-to-dialogue transformer generating conversational speech with optional voice cloning.
An intermediary that translates AWS Bedrock API calls into the Anthropic API format so existing clients can use Bedrock.
MCP server with tools to search, retrieve, and navigate municipal ordinances and codes from Municode's digital library.
Food tour planner using LangChain DeepAgents, Google Maps, and Tavily to plan personalized neighborhood food tours.
AI agent skill researching any topic across Reddit, X, YouTube, HN, Polymarket, and the web, scored by engagement.
A Next.js portfolio with reusable Tailwind and shadcn/ui components and MDX-powered content.
Wayland compositor with scrollable tiling windows on infinite horizontal strips, dynamic workspaces, and multi-monitor support.
A paper on olmOCR-2 describing unit-test-based reward training for improving document OCR quality.
Data recipes and infrastructure for training AI agents, providing datasets and tooling for developing small agentic models.
A Codex CLI release with a minor fix to a capacity error message, following the larger 0.15.0 upgrade.
A Claude Code plugin to run Codex code reviews and delegate programming tasks without leaving the workflow.
An open-source framework that turns project work into isolated, autonomous coding-agent runs so teams manage work instead of supervising agents.
An open-source framework for building conversational agents with multi-channel integration and model management.
Open-source RL framework letting LLMs learn from experience via GRPO, supporting Qwen, Llama, and serverless infrastructure.
A package in the opral monorepo, home of inlang, the open-format translation management system for software teams.
A persistent multi-agent development environment on Claude Code that keeps context across sessions via state management and specialized agents.
Benchmarks language models on the NYT Connections puzzle, comparing baseline vs DSPy-optimized performance with MLflow tracking.
A SQLite-based filesystem for AI agents offering auditability, reproducibility, and portability via structured state and operation history.
A reproducible evaluation framework for testing deep-research web search systems across challenging benchmarks.
A Python framework for building environments with datasets, harnesses, and reward functions to train and evaluate LLMs via RL.
A community-curated directory of publicly available free APIs across many categories for use in development projects.
A sandboxed Python interpreter in Rust for safely running LLM-generated code, with fast startup, resource limits, and no default I/O access.
A desktop app that uses various LLMs to build and run full-stack applications locally on your computer.
A terminal-based AI agent optimized for the Qwen model series to understand codebases and automate coding tasks.
An enterprise-grade RL framework for large-scale language and vision model post-training with FP8 training, speculative decoding, and stability.
A curated list of ports, filesystem adapters, libraries, and integrations for just-bash, a virtual bash environment for agents.
A tool using AI to perform automated transformations on code repositories, such as porting projects across languages and frameworks.
Lightweight in-process runtime letting AI agents start in ~6ms at far lower cost than traditional sandboxes.
An npm package that runs untrusted JavaScript securely using V8 isolates, with no containers, VMs, or infrastructure required.
A drag-and-drop tool for creating agentic workflows with interactive debugging and TypeScript code generation.
An automated auditing tool that tests language models for alignment issues via multi-turn auditor and judge scenarios.
Framework partitioning transformers into independently trainable blocks to cut memory while keeping performance.
A browser automation tool letting AI agents control browsers via sandboxed JavaScript scripts with persistent pages and full Playwright API support.
A local-first platform that orchestrates many AI coding agents under one manager, dispatching workers in parallel and surfacing only key decisions.
JS system letting multiple coding agents collaborate through deterministic parallel and sequential workflows.
Open-source tool for iteratively improving AI agent prompts through sampling, evaluation, and optimization.
A lightweight LLM inference framework in ~5,000 lines of readable Python featuring radix cache, tensor parallelism, and FlashAttention.
Agent skill that audits codebases and generates executable implementation plans for cheaper models to run.
An open-source AI workspace to build, deploy, and manage AI agents via conversational, visual, or code interfaces.
Autonomous AI agent that iteratively completes PRD items using fresh context windows, with memory persisted via git history.
An MCP server enabling AI assistants to interact with Jira and Confluence across Cloud and Server/Data Center deployments.
A toolkit that runs automated evaluations testing how well AI models handle Next.js development tasks.
Local-first encryption and dashboard for debugging AI agent transcripts from Codex and Claude without remote services.
DSPy source implementing GEPA, an evolutionary optimizer that uses reflection to evolve text components of modules.
A 'tinker' skill module within the sundial-org/skills repository providing developer-focused educational content.
Agent Skills to help developers use AI agents with Supabase, compatible with 18+ agents including Claude Code and Copilot.
Minimal Python AI agent that solves software tasks via LLM plus bash, scoring over 74% on SWE-bench while staying radically simple.
MCP server giving agents access to your architectural decisions, requirements, and designs mid-task.
Open-source monorepo providing observability and instrumentation for AI agents across TypeScript and Python with local-first UI.
An AI agent framework that uses TypeScript code execution as its primary tool to interact with APIs and complete tasks.
A Rust tool for semantic search and natural-language queries over codebases using embeddings and tree-sitter parsing.
An on-device search engine combining BM25, vector search, and LLM re-ranking to index and query markdown documents locally.
Repository holding the formal spec for Token-Oriented Object Notation (TOON), a line-oriented encoding format for JSON data.
Connects Claude Code to local TradingView Desktop via CDP for AI-assisted chart analysis and Pine Script development.
A framework letting language models recursively call sub-models with DSPy signatures for interpretable trajectories.
Claude skill applying Edward Tufte's data-viz principles to transform cluttered charts into clean, minimal designs.
Processes large data collections with LLMs via map/reduce/filter operations while optimizing pipelines for accuracy and cost.
LLM-powered platform for autonomous discovery, backtesting, and live execution of strategies on Polymarket and Kalshi.
Transforms web content into clean, AI-ready Markdown through intelligent extraction and structuring.
Hierarchical document indexing that uses LLM reasoning for retrieval instead of vector similarity, removing chunking and vector DBs.
Fast native Rust CLI automating browser interactions via Chrome DevTools Protocol, built for AI agent workflows.
A template for AI coding agents supporting Claude, OpenAI, Copilot, Cursor, and Gemini that run tasks in isolated sandboxes.
Web-based terminal emulator with DOM rendering written in Zig compiled to WASM, with native text selection and a11y.
Platform for searching, inferring over, and organizing vectors, tensors, text, and structured data at scale and serving time.
A desktop app that transcribes meetings, extracts action items, and generates AI summaries using OpenAI's Realtime API.
Turns any project into a tmux-based terminal IDE via a simple ide.yml defining pane layouts and commands.
A Python collection of pre-built compound AI retrieval systems combining DSPy and Weaviate.
GitHub Actions workflow that uses Claude to analyze git commits and auto-generate customer-facing weekly changelogs.
TypeScript framework for building autonomous agents that maintain context, use tools, and safely take action in sandboxed environments.
Lets Copilot CLI query Microsoft 365 data including emails, meetings, documents, Teams messages, and people for contextual insights.
Converts docs sites, repos, PDFs, videos, and more into structured knowledge assets for Claude, Gemini, OpenAI, and RAG pipelines.
A Chrome extension turning the new tab page into a dashboard to organize open tabs by domain with duplicate detection.
Python framework using reinforcement learning to optimize and fine-tune DSPy language model programs at scale.
A curated collection of 50 papers and resources across 10 AI domains for engineers building frontier LLM knowledge.
Guide covering Claude Code 2.0 features, workflows, and context engineering for using AI coding agents.
An OpenAI business guide covering practical approaches and patterns for building AI agents.
Introduces mcp-to-ai-sdk, a CLI that generates static, vendored tool definitions from MCP servers for safer production AI agents.
Tool evaluating how well products, SDKs, and docs work when agents attempt real onboarding from minimal prompts.
Course on building multi-agent systems with Google's ADK to construct knowledge graphs from structured and unstructured data.
Deep Agents combine explicit planning, hierarchical delegation, persistent memory, and prompting to handle complex multi-step tasks.
A course teaching how to build, evaluate, and improve AI agents through systematic evaluation and error analysis.
Documentation for building AI-powered applications with Vercel's framework, supporting multiple providers and React hooks.
Collection of 133 copy-paste patterns for building AI agents with Vercel's AI SDK v6, covering tool calling and orchestration.
Guide to using the AI SDK within coding agents like Claude Code and Cursor, covering skills, local docs, and DevTools.
A registry for installing AI SDK tools and components via the shadcn CLI, with UI renderers for chat result displays.
A lightweight runtime that runs Node.js, Next.js, Vite, and Express apps entirely in the browser without a backend server.
S3 Files provides high-performance file system access to data in Amazon S3 using standard file operations without duplication.
A Kaggle notebook using autoencoders to detect anomalies by reconstructing normal data and flagging instances that deviate significantly.
An AI app builder that turns natural language into working mobile apps, websites, and tools with 40+ integrations.
A storage solution and cloud filesystem designed specifically for AI applications and workflows.
Documentation for the disk package, a JavaScript control-plane client for managing Archil disks and running serverless commands from Node.js.
Enables agents to register users without a sign-up form, providing secure agent registration any app can add.
OpenAI's Auto-review mode for Codex uses a separate agent to approve boundary-crossing actions, cutting approval interruptions 200x.
A tutorial on customizing DSPy to automatically generate and optimize LLM system prompts using adapters and optimizers.
A framework that compiles typed function signatures into reliable LLM calls with validation, streaming, tools, and agents.
An observability platform for ingesting, storing, and analyzing logs, metrics, traces, and events at petabyte scale.
Reference documentation describing quotas, rate limits, and quota tiers for Azure OpenAI models in Microsoft Foundry.
Anthropic's agentic framework that auto-generates scenarios to measure how often misaligned behaviors occur in frontier AI models.
Instant Linux VMs with pre-installed dev tools built for AI agents, with affordable per-second billing.
A tool that makes websites accessible to AI agents by extracting interactive elements so agents can control browsers via natural language.
A command-line interface for fast browser automation, controlling websites and managing cloud browser platforms from the terminal.
A read-only CDP client that records browser automation traces to NDJSON, capturing screenshots and DOM snapshots.
Developer community offering free credits, Discord access, build challenges, and XP rewards for shipping AI-agent projects.
A filesystem-based approach to knowledge agents that replaces vector databases with bash commands for cheaper, more transparent retrieval.
Webinar by Prof. Tom Yeh on creating a vector database using Excel, with downloadable resources and a recording.
Developer builds a neural-embedding search engine over 280M web pages using 200 GPUs, achieving 500ms query latency.
Argues bash and filesystem tools replace brittle custom tooling, giving LLMs a simpler, more effective agent architecture.
Technical guide on using Cloudflare Durable Objects to create per-tenant graph databases with SQLite storage.
Lee Robinson builds pixo, a zero-dependency Rust image compression library, entirely with AI coding agents without writing code by hand.
Reflection on the challenges of using AI agents to build Pi, including low-quality bug reports and over-engineered contributions.
Docs for the security-guidance plugin that makes Claude review its own code changes for vulnerabilities and fix them in-session.
A testing and monitoring platform to evaluate, simulate, and continuously improve conversational AI agents before and after deployment.
Decentralized platform providing serverless compute for deploying, scaling, and running open-source AI models in production.
Guide to building high-accuracy classifiers with Claude using prompt engineering, RAG, and chain-of-thought reasoning.
An experiment comparing zero-shot and few-shot prompting with Gemini Pro for customer ticket classification, finding few-shot most accurate.
Anthropic's technical specifications, capabilities, and safety considerations for the Claude 4.1 model.
A guide on security principles for AI agents, covering verification, tool access controls, and emerging threats.
Explains how agents interact with MCP servers more efficiently by writing code instead of direct tool calls to cut token use.
Cloudflare's MCP server uses Code Mode to expose 2,500+ API endpoints through two tools using ~1,000 tokens instead of over a million.
Cloudflare proposes converting MCP tools into TypeScript APIs for LLMs to write code against, improving agent performance.
An introductory tutorial on AI evaluation methodologies by Aman Khan, designed to be completed in about an hour.
A macOS app to run multiple coding agents (Claude Code, Codex, Cursor) concurrently in isolated workspaces.
Use Claude Code as a testing harness to validate agent concepts in hours by exposing tools as CLI commands before building infrastructure.
Argues AI agents should learn through context-token updates rather than model weights for more portable, controllable learning.
An enterprise platform connecting AI agents to UIs across web, Slack, and Teams using the AG-UI protocol for generative UI.
Anthropic and CMU research on Incalmo, a toolkit enabling LLMs to execute multi-stage cyberattacks on realistic networks.
A talk on developer tools like the Apps SDK, Agent Kit, and MCP, stressing the growing importance of effective prompting.
OpenAI's developer notes on updates and capabilities of the Realtime API, including the new gpt-realtime speech-to-speech model.
Wasmer open-sourced Edge.js, a runtime for executing Node.js apps in WebAssembly sandboxes with fast startup and full Node compatibility.
Lets AI agents create videos by writing declarative HTML and CSS compositions that render across browser, CLI, and cloud.
Amp now lazy-loads MCP server tools via skills, cutting token usage from 17k to 1.5k for chrome-devtools.
CLI that captures AI agent sessions alongside git commits, creating searchable checkpoints preserving the reasoning behind changes.
Anthropic introduces Agent Skills, organized folders of instructions and resources that let Claude agents specialize via progressive disclosure.
A TypeScript testing framework for evaluating AI apps with a local dev server, interactive UI, and Vitest integration.
Factory.ai tested context compression methods and found structured summarization preserves more useful info than alternatives.
Guide to assessing Q&A systems on long documents, covering metrics like faithfulness and helpfulness, methods, and benchmarks.
Survey of how LLMs evaluate other LLM outputs, covering prompting, alignment, finetuning, and practical considerations.
Cloud platform offering disposable VMs, persistent VPS, and dev environments for developers and AI agents with SSH and root access.
Explores how AI agents increasingly use files as a core abstraction for managing context and executing tasks over tool-calling.
Vercel Sandboxes now capture and reuse filesystem states via snapshots, eliminating repetitive setup for costly operations.
A DSPy tutorial on optimizing chain-of-thought reasoning for competition math problems using the GEPA optimizer.
An Observable notebook exploring how iterative GEPA prompt refinement can outperform reinforcement learning.
Guide to connecting the Atlassian Rovo MCP Server (Jira, Confluence, Compass) to external tools via OAuth 2.1.
Gives AI agents a governed graph to explore databases, files, APIs, and code through a single GraphQL query with access boundaries.
Examines the math behind RL methods for LLMs, comparing GRPO with its variants Dr.GRPO and GSPO.
A GAN-inspired multi-agent architecture using generator and evaluator agents to build full-stack apps autonomously.
Instant web hosting and private cloud storage letting AI agents publish files, sites, and apps to live URLs.
Decagon and Modal partnered to build fast voice AI agents via model fine-tuning and optimized inference.
Explains LLM response generation: tokenization, transformer architecture, prefill and decode phases, KV caching, and production serving.
Explains how LLM providers implement prompt caching via paged attention and block hashing to reuse computed KV tensors across requests.
Reducto explains its ML approach to accurately extract numerical data from document charts using multiple lightweight component-detection models.
A guide to building a functional code-editing agent using Claude, the Anthropic API, and Go in about 300 lines of code.
Hamel Husain and Emil Sedgh discuss building evaluation systems tailored to specific domains for large language models.
A comprehensive guide to the Model Context Protocol and building production-ready remote MCP servers with auth and session management.
Design skill giving agents and developers shared vocabulary for polished interfaces, eliminating common AI-generated design patterns.
An open-source framework for large language model evaluations across coding, reasoning, and knowledge capabilities.
Vercel open-sources bash-tool, a Bash execution engine letting AI agents retrieve context by running filesystem commands instead of loading large prompts.
LangChain's Deep Agents CLI lets developers build agents from the terminal with persistent memory, file management, and web search.
Vercel open sourced deepsec, an AI-powered tool that finds hard-to-detect vulnerabilities in large codebases using Claude and agents.
Databricks open-sourced Omnigent, a platform unifying multiple AI agents across harnesses and models.
A unified TypeScript library that lets teams write bot logic once and deploy it to Slack, Teams, Google Chat, Discord, GitHub, and Linear.
A skill package for structured AI-assisted development that stores requirements, plans, and decisions in repo files, not chat.
A proof that Jira achieves computational universality by implementing a Minsky machine via automation rules.
A browser infrastructure platform enabling web agents and automations to navigate and interact with web environments.
Introduces Kimi's text and multimodal models, tokens, rate limits, API keys, and the Chat Completions API.
Toolchain enabling AI agents to autonomously build and deploy full-stack applications without complex infrastructure.
A course on building reliable AI coding agents through environment design, state management, and verification.
LLMs blur the programmer/non-programmer line, shifting focus to safe sandboxes for executing user-specific software.
A personal AI assistant that manages inbox, meetings, calendar, scheduling, follow-ups, and CRM across apps.
Tool that continuously replicates SQLite databases to cloud storage, enabling safe single-server deployment with disaster recovery.
Open source framework and cloud platform for creating voice, video, and physical AI agents.
A benchmark evaluating how well language models discover and use Model Context Protocol tools across diverse domains.
A tool that streams Effect spans in real time to React applications.
A comprehensive FAQ covering LLM evaluation fundamentals, error analysis, human annotation, tooling, and production deployment.
A comprehensive FAQ covering LLM evaluation fundamentals, error analysis, human annotation, tooling, and production deployment.
Autonomous LLM-powered trading bot that monitors social media for trending stocks, analyzes sentiment, and executes trades safely.
A free service that converts any public URL into clean Markdown optimized for AI agents, cutting token usage by roughly 80% versus raw HTML.
Introduces SEP-1865, a standardized extension letting MCP servers deliver interactive UIs, built with OpenAI, Anthropic, and the MCP-UI community.
Platform for unified integration, governance, and monitoring of AI agents connecting tools like Salesforce, Slack, and Jira.
Virtualization layer letting AI agents access multiple backends through one virtual filesystem using bash commands and Unix-like pipes.
Presents five multi-agent coordination patterns with guidance on choosing the right one for a task.
An open-source platform for building and running background coding agents on Vercel with web UI and sandbox orchestration.
Examines five real-world implementations of OpenAI's Realtime API across industries, showing latency gains and business outcomes.
Open-source local-first AI assistant that manages tasks via WhatsApp, Telegram, and other chat apps with persistent memory.
MLflow's prompt optimization systematically enhances AI apps using GEPA or Metaprompting with minimal code and no framework lock-in.
Comparing two candidate outputs for the same input and picking the better one, often with an LLM judge.
A tool that converts web pages into organized, machine-readable APIs for structured data extraction.
Reverse-engineers Claude Code's API calls, finding its strength comes from prompt scaffolding and sub-agent architecture.
Minimal, extensible agent harness that adapts to workflows via customizable extensions, skills, and 15+ AI provider integrations.
A practical methodology for building product evaluations with LLM-based assessors using binary labeling, evaluator alignment, and automated testing.
Open-source GitHub bot that runs AI agents to automate code reviews, issue triage, and CI fixes inside GitHub Actions.
Databricks' Genie data agent improves accuracy from 32% to over 90% using knowledge search, parallel thinking, and multi-LLM methods.
Introduces PydanticAI, a type-safe, model-agnostic Python framework from the Pydantic team for building production-grade AI agents.
A tutorial covering PyTorch tensors, autograd, neural networks, data loading, training loops, and multi-GPU training.
Shopify's Quick platform lets employees instantly deploy websites by uploading an HTML folder.
Platform that traces AI agent execution, automatically detects failures, and helps teams debug via Slack workflows and experiments.
Language models that recursively call themselves through REPL environments to handle unbounded context lengths.
A DSPy tutorial introducing GEPA, an optimizer that reflects on program trajectories to propose improved prompts.
Interactive library showcasing completions from instruction-tuned and RLHF-trained models like OLMo and Tulu for educational use.
The official digital gateway for South Carolina state government services for citizens, businesses, and visitors.
Cookbook showing how to build feedback loops that let AI agents autonomously improve via iterative evaluation and prompt optimization.
A spreadsheet agent that thinks and acts for you to automate spreadsheet tasks with AI.
A leaderboard benchmarking AI model-agent configurations on the SkillsBench suite, comparing resolution rates with and without specialized skills.
A lightweight native terminal workspace that runs AI coding agents alongside your dev stack with process monitoring and MCP integration.
The SC agency providing behavioral health, developmental disabilities, and substance use services.
Documentation for Claude Code fast mode, a high-speed Opus configuration toggled with /fast that trades higher cost for up to 2.5x faster responses.
Argues auto-generated MCP servers hurt agent performance; instead curate APIs designed for LLMs rather than humans.
Researchers found a critical MCP vulnerability enabling arbitrary command execution, affecting 200+ open source projects.
An Excel exercise demonstrating how temperature influences token sampling and probability distributions in language models.
MCP update introducing a stateless core architecture, extensions framework, hardened authorization, and a formal deprecation policy.
Explains how a harness wraps an AI model to maintain state, run tools, and complete autonomous tasks.
An Anthropic guide on designing and building Skills for Claude.
Zero-configuration functions for building AI apps with streaming, tool-calling, and structured output across providers.
A guide on using Claude to classify and route customer support tickets at scale by intent, urgency, and other factors.
Anthropic docs explaining how Claude calls client and server tools, where they execute, and how the agentic loop works.
AWS docs explaining how to implement tool use, RAG, and agentic workflows with the Amazon Nova Sonic speech model.
Platform for building and deploying TypeScript AI workflows with task management, observability, and elastic scaling.
Cloud-native SQLite database platform for millions of lightweight databases with concurrent writes, vector search, and AI support.
Shopify built Aquifer, a durable agent platform under River, a Slack-native AI coding assistant for shipping agents at scale.
Covers the experimental uv format command in uv 0.8.13, which integrates Ruff's formatter into uv.
Independent platform that benchmarks leading AI models on domain-specific tasks across finance, law, healthcare, and software.
Platform for building, training, and deploying AI models with experiment tracking, hyperparameter tuning, and monitoring.
Explains MCP prompts as reusable, parameterized message templates from servers that guide interactions with agents.
Argues that agents running Bash and writing composable code are often more efficient than using MCP servers.
Vercel shares its method for picking high-impact AI agent projects, focusing on repetitive, low-cognitive-load tasks.
Google announces a lawsuit against SerpApi for circumventing security to scrape and resell copyrighted Google Search results.
An open-source Next.js template for building a visual automation platform with a drag-and-drop editor, AI generation, and prebuilt integrations.
Guidance on using coding agents effectively, covering harnesses, prompt strategies, context management, and scope control.
Tutorial teaching how to build a VM that executes LC-3 assembly programs in about 250 lines of C.
Pricing page for YouMind, an AI-powered tool offering prompts and tools across generative models for different user tiers.
Independent evaluation tools for AI systems, starting with LLM Stats, a benchmarking platform for high-stakes applications.
Framework for training LLM agents via RL with step-level trajectories and flexible context for multi-round tool use.
A framework that treats prompts as evolving playbooks that accumulate and refine strategies for more effective LLM agents.
Introduces a conformal-prediction framework to quantify uncertainty in LLM-based evaluation via score prediction intervals.
Reframes the ARC benchmark as image-to-image translation, using Vision Transformers to reach 60.4% on ARC-1.
An evaluation suite of ~800 original problems across seven scientific fields to test frontier LLM cross-disciplinary reasoning.
A meta-agent programming approach that automatically discovers novel agentic system designs by iteratively coding better agents.
Introduces a Budget Tracker and BATS framework so tool-using agents strategically decide when to dig deeper versus pivot based on remaining budget.
Listing of recent research papers in computational linguistics and NLP from arXiv's cs.CL category.
Listing of the latest research papers submitted to arXiv's Computer Vision (cs.CV) category.
Technical report analyzing Claude Code's architecture to identify design principles for agentic systems.
A study finding that repository-level context files tend to reduce coding agent task success while increasing inference cost by over 20%.
Examines core principles and techniques for developing autonomous LLM-powered agents, covering architecture and implementation.
Explains how transformer blocks let LLMs learn new patterns at inference via implicit MLP-layer weight modifications.
Introduces a benchmark of 895 synthetic questions with difficulty scores to evaluate Retrieval Augmented Generation systems.
A survey examining how large language models can act as evaluators across functionality, methodologies, applications, and limitations.
Listing of the latest research papers submitted to arXiv's Machine Learning (cs.LG) category.
A vision-language-action model that improves via real-world RL to perform complex robotic tasks like laundry folding.
A method letting LLMs process prompts beyond their context window by recursively decomposing input snippets.
A method for language models to recursively decompose complex problems into subproblems, improving performance on long-horizon reasoning tasks.
Proposes an efficient spatiotemporal framework that exploits data sparsity to cut compute costs while preserving performance.
Presents SWE-Gym, an environment of 2,438 real-world Python tasks for training AI agents, with SOTA software engineering results.
A unified multimodal model improving image understanding, generation, and editing via a shared RL reward strategy.
Explains how continuous batching with KV caching, chunked prefill, and dynamic scheduling maximizes throughput in LLM serving.
A 21B-parameter MoE text model (3B activated) with enhanced reasoning, tool use, and 128K context length.
A 357B-parameter model with a 200K context window featuring enhanced coding, reasoning, and tool use for agentic tasks.
Guide to the hf CLI for logging in, creating repos, and uploading/downloading files on the Hugging Face Hub.
A 32B open-weights reasoning model with strong performance in math, code generation, and scientific tasks.
A TRL feature shrinking async RL weight sync to sparse deltas (99% smaller) for disaggregated distributed training.
A Hugging Face Space providing resources and guidance for training smaller, efficient language models.