Agent Community, Directories & Research

Quora's Poe platform allows developers to deploy custom AI bots built on any LLM with a massive distribution channel. Bots reach millions of Poe users without building your own frontend. Monetization through Poe's creator program.

Platform for hosting and discovering ML demos and AI agent applications built with Gradio, Streamlit, or Docker. 500,000+ apps hosted publicly, making it the largest directory of interactive AI demos. Free GPU quota for community apps.

Official registry of LlamaIndex data loaders, tools, and agent integrations contributed by the community. 300+ connectors for loading data from any source into LlamaIndex-powered agents. One-line install with the LlamaIndex CLI.

Community cookbook of production RAG and agent patterns built on Weaviate. Covers hybrid search, multi-tenancy, agents, and fine-tuning integration. The fastest way to learn real-world vector database patterns from working code.

Curated GitHub list of AI agent frameworks, platforms, and products maintained by the E2B team. Organized by framework, product type, and capability. One of the most-starred references for navigating the AI agent ecosystem.

Hugging Face's automated benchmark leaderboard for open-weight models across academic tasks. Evaluates models on ARC, HellaSwag, MMLU, TruthfulQA, and newer benchmarks. The primary reference for comparing open-source model capabilities.

Crowdsourced LLM benchmarking platform where users compare model outputs head-to-head in blind A/B tests. The Elo-based leaderboard is considered one of the most reliable measures of real-world model quality. Tens of millions of human preference votes.

Community hub for sharing and discovering Superagent-based AI assistants and agent configurations. Browse pre-built agents for sales, support, and research use cases. Fork and deploy community agents directly to the Superagent cloud.

Official collection of reference applications and production-ready LangChain templates. Covers RAG, extraction, chatbots, agents, and more — each deployable with a single command. The fastest way to bootstrap a production LangChain application.

Transformers Agents is a multi-modal agent API built into the Hugging Face ecosystem. Provides a natural language interface for calling 100,000+ HF models as tools. Integrates with the Hub model, dataset, and Space ecosystem for end-to-end AI agent workflows.

Open-source LLM application development platform that combines BaaS and LLMOps. Provides a visual workflow builder, RAG pipeline, agent framework, and model management in one product. One of the fastest-growing open-source AI application platforms.

French-origin multilingual AI tools directory with 5,000+ tools in 15+ languages. Includes user reviews, ratings, and categorized listings across 50+ use case categories. Strong coverage of European and international AI tools.

Daily AI newsletter with 100,000+ subscribers covering the latest tools, research, and industry moves. Written by Ben Tossell, the founder of Makerpad. The most consistently useful daily brief for developers and founders building in AI.

Largest AI tools directory aggregating thousands of AI tools across categories.

Product Hunt's AI category featuring the newest AI tools launched by founders. Daily featured products include agent frameworks, AI APIs, and automation tools. Community upvoting surfaces genuinely useful tools. Most significant new AI products launch here first.

AI tools discovery platform with 30,000+ tools organized by category, revenue, and traffic rankings. Tracks monthly visits to identify the most popular AI products. Useful for competitive analysis and discovering which agent tools are gaining real-world adoption.

Curated collection of AI tools with filters, news, and community ratings.

Curated collection of the best AI tools by The Rundown AI newsletter. Expert-vetted with descriptions of real use cases. Newsletter with 700K+ subscribers drives significant discovery traffic. Focuses on practical, production-ready tools rather than experimental projects.

Clean, fast directory of AI tools with weekly additions and community ratings. Organized by category with filtering by pricing tier (free, freemium, paid). Simple design optimized for discovery. Useful for finding affordable or free alternatives to commercial AI agent tools.

Community repository for sharing and discovering LangChain prompts, chains, and agent configurations. Version-controlled prompt management with pull-from-hub functionality. The largest collection of production-tested LangChain prompts contributed by the community.

The largest AI tools directory aggregating thousands of tools across all categories. Useful reference for tracking what already exists in the broader AI space.

OpenAI's official marketplace for custom ChatGPT agents (GPTs). Browse thousands of community-built specialized agents for coding, research, creative writing, and analysis — or publish your own GPT to reach millions of ChatGPT users without building a standalone product.

Community platform for discovering, sharing, and building browser-based AI agents. Provides a visual agent builder, a catalog of pre-built agents for common tasks, and tools for publishing agents as shareable URLs — democratizing agent creation for non-engineers.

Curated directory of AI agents organized by category and use case. Updated regularly with new agent products and open-source projects. Includes reviews, GitHub links, and deployment information for each listed agent. Good starting point for discovering production-ready agent solutions.

Create and share custom AI assistants powered by open-source LLMs on Hugging Face Hub. Enables developers and researchers to publish specialized agents built on models like Llama, Mistral, and Falcon — a key distribution channel for the open-source AI community.

Largest aggregator of AI tools with 50,000+ tools across 3,000+ categories and 8M+ monthly visits. Features trending tools, new additions, and use-case search. A go-to resource for discovering AI agents and tools in any vertical.

Open-source drag-and-drop UI for building and deploying AI agents and LLM workflows. Provides a visual canvas for wiring together LLMs, vector stores, tools, and memory components — making it possible to build production agents without writing a single line of code.

Curated newsletter and directory of AI tools maintained by Matt Wolfe. Weekly AI news alongside a searchable database of AI tools with community ratings and categories. Large YouTube following ensures consistent discovery by new AI practitioners.

Open API specification standardizing how AI agents expose their capabilities and accept tasks. Enables interoperability between agent frameworks, benchmarking tools, and deployment platforms. Growing adoption in CrewAI, AutoGPT, and other major frameworks.

Visual framework for building and deploying multi-agent AI pipelines. Offers a low-code editor for composing complex agent architectures, with support for LangChain components, custom Python code execution, and one-click deployment to cloud platforms.

Platform hosting 500,000+ AI demos and applications. The largest collection of publicly deployable ML models and agent demos. Gradio and Streamlit apps run free on shared CPU/GPU. The de facto standard for sharing open-source AI research prototypes.

Template marketplace for Flowise agentic workflows. Pre-built agent templates for customer support, data analysis, code generation, and document Q&A. Templates can be imported with one click and customized for specific use cases. Growing community contribution.

Multimodal agent framework for operating smartphone apps by analyzing screenshots and generating touch actions. Does not require access to app source code or view hierarchies. Shows strong generalization to novel apps after a short exploration phase.

DeepMind's system ranking in the top 15% of competitive programmers on Codeforces. Combines Gemini models with a specialized fine-tuning and sampling strategy. Demonstrates superhuman coding agent capabilities on algorithmic competition tasks.

Benchmark for evaluating web-based AI assistants on realistic time-consuming tasks that require planning and multi-step web navigation. 214 tasks across information retrieval, data aggregation, and transaction tasks. Current top agents score under 25%.

LLM benchmark created from real-world user interactions rather than curated academic tasks. 1,024 challenging tasks collected from ChatBot Arena with automated scoring using GPT-4. More representative of actual user needs than traditional benchmarks.

The evolution of LMSYS Chatbot Arena with expanded multimodal capabilities and model coverage. Battle-based human evaluation now covering vision and coding capabilities. Tens of millions of preference votes making it the most reliable human-preference benchmark.

Microsoft's prompting strategy that achieved state-of-the-art results on medical benchmarks with GPT-4, outperforming specialized medical models. Combines dynamic few-shot selection, self-generated chain-of-thought, and choice shuffling. Shows that prompting alone can match fine-tuning.

Comprehensive benchmark for evaluating LLM-based agents across 8 distinct environments including code, databases, web browsing, and games. Reveals major performance gaps between open and closed models on real-world agentic tasks. The de-facto standard for comparing agent capabilities.

OpenAI's foundational paper on training GPT to answer questions using a web browser. Demonstrated that RLHF can teach LLMs to search, navigate, and cite web pages. Laid the groundwork for all subsequent web-browsing agents.

Research framework and dataset for teaching LLMs to use 16,000+ real-world REST APIs via tool calling. Introduces the DFSDT algorithm for efficient API exploration and a benchmark for measuring tool-use proficiency.

Paper showing that LLMs can iteratively improve their own outputs using multi-step feedback loops without additional training. Influential technique for building self-correcting agents. Applicable to code generation, essay writing, and mathematical reasoning.

Google Brain's landmark paper demonstrating that prompting LLMs to produce intermediate reasoning steps dramatically improves performance on complex tasks. Foundational technique underlying almost all modern agent reasoning strategies. 10,000+ citations.

System that uses ChatGPT as a task planner to coordinate specialized Hugging Face models for multimodal tasks. Demonstrates an "agent as orchestrator" pattern where the LLM selects and invokes the right specialized model for each sub-task.

Comprehensive survey of 600+ papers on LLM-based agents covering architecture, training, application, and evaluation. Provides a unified taxonomy for understanding the agent design space. The most cited review paper in the field.

Paper showing that asking LLMs to first generate a plan before solving a problem significantly improves multi-step reasoning. Introduces "Plan-and-Solve+" with a step-level trigger. Widely implemented in agent planners as a cheap, training-free improvement.

Meta's paper on retrieval-augmented language model pre-training. Shows that a small 11B parameter model with retrieval can match GPT-3's 175B parameter performance on knowledge-intensive tasks. Foundational work for understanding why RAG is so effective.

OpenAI's technical report on o1 — a model trained with reinforcement learning to reason before answering. Demonstrates that extended chain-of-thought at inference time substantially improves performance on hard math, science, and coding problems. A new paradigm for agent reasoning.

DSPy's automated prompt optimization algorithm that outperforms human-written prompts on complex tasks. Uses Bayesian optimization over instruction and few-shot example combinations. Demonstrates that prompt engineering can be automated with surprising effectiveness.

Vision-language model agent that completes real-world web tasks by processing screenshots. Demonstrates that visual grounding (seeing the page) outperforms HTML-based agents on diverse navigation tasks. Strong baseline for multimodal web agents.

First dataset and benchmark for generalist web agents capable of following language instructions to complete tasks on 137 real websites. 2,000 tasks spanning travel, shopping, and information retrieval. The standard benchmark for web navigation agents.

Seminal paper combining reasoning traces and acting in language models for agent tasks.

Primary online venue for AI safety and alignment research. Hosts technical discussions on agent goals, corrigibility, scalable oversight, and related topics. Many foundational ideas for safe agentic AI systems originated here. Essential reading for building agents responsibly.

The preprint server where most AI research appears before peer review. The cs.AI and cs.LG sections publish hundreds of papers weekly on agent architectures, benchmarks, and techniques. Subscribe to daily digests to stay current with the latest academic advances.

Browse autonomous agent research papers alongside their official code implementations. State-of-the-art leaderboards for agent benchmarks updated in real-time. Filter by task (code generation, web navigation, embodied AI) to find the latest best-performing methods.

Open-ended embodied agent that continuously acquires new skills in Minecraft.

Crowdsourced benchmark for LLM evaluation through blind A/B tests. Over 1 million human comparisons powering the Elo-based leaderboard. Identifies which models perform best on instruction following and reasoning — critical data for choosing the right LLM backend for agent applications.

Free open-source two-volume MIT Press textbook on Machine Learning Systems covering training, inference, edge deployment, distributed systems, and AI governance. 25k+ GitHub stars.

Build your own ML framework from scratch across 20 progressive modules, part of the Harvard ML Systems curriculum. The best way to deeply understand how frameworks like PyTorch work internally.

Physics-grounded ML systems interview practice with vault, drills, mock interviews, and progress tracking. Covers cloud, edge, mobile, and TinyML system design questions.

Stanford's landmark paper simulating believable human behavior in a virtual village with 25 LLM-powered agents. Demonstrates emergent social behaviors including relationship formation, memory consolidation, and daily routine planning — a seminal work on multi-agent social simulation.

Seminal 2022 paper introducing the ReAct framework, which interleaves reasoning traces and task-specific actions in LLMs. Demonstrates superior performance and interpretability over chain-of-thought and action-only approaches. The theoretical foundation for most agent reasoning patterns in use today.

Facebook Research paper showing LLMs can teach themselves when and how to use external tools via self-supervised learning. Models generate their own training data for tool use by trying tools and keeping calls that improve loss. Foundational work for tool-using agents.

Meta's research paper demonstrating that LLMs can learn to use external tools (calculators, search, translation APIs) by self-supervising on when and how to make API calls. Showed that tool use can be taught through fine-tuning without explicit human annotation of tool invocations.

Stanford and Google Research paper creating interactive simulacra of human behavior. 25 LLM-powered agents live in a sandbox, form memories, plan daily activities, and interact socially. Demonstrates emergent social behaviors from individual agent architectures.

Introduces the Reflexion framework for agents that improve performance through verbal self-reflection. Instead of updating weights, agents verbally analyze failed attempts and store those reflections as memory — achieving significant performance gains on coding, reasoning, and decision-making benchmarks without any gradient updates.

NVIDIA and Caltech paper presenting the first LLM-powered lifelong learning agent in Minecraft. Automatically discovers and masters new skills, building a growing library of reusable knowledge. Demonstrates continual skill acquisition without human intervention using GPT-4.

Proposes a framework for deliberate problem solving with LLMs by exploring multiple reasoning branches simultaneously. Enables agents to evaluate different solution paths, backtrack from dead ends, and perform systematic search over the space of possible thoughts — dramatically improving performance on complex multi-step problems.

Documents the architecture and empirical analysis of AutoGPT, one of the first widely-deployed autonomous agent systems. Analyzes the task completion success rates, failure modes, and emergent behaviors observed across thousands of real user deployments — providing the field with its first large-scale empirical study of autonomous agents in the wild.

Comprehensive 2023 survey paper covering the architecture, capabilities, and applications of LLM-based autonomous agents. Reviews 150+ papers across planning, memory, tool use, and multi-agent systems. Essential reading for understanding the full state of the field.

Iterative self-improvement paper showing LLMs can critique and refine their own outputs without additional training. The agent generates output, provides feedback on it, and refines based on that feedback in a loop. Applicable to code generation, math, and creative writing tasks.

Princeton and Google DeepMind paper extending chain-of-thought to tree-structured exploration. LLMs explore multiple reasoning paths simultaneously and use search algorithms (BFS/DFS) to find optimal solutions. Dramatically improves performance on planning and multi-step reasoning tasks.

ETH Zürich extension of Tree of Thoughts enabling non-linear thought structures. Thoughts can be combined, merged, and looped in arbitrary graph topologies. Enables solving problems that require aggregating insights from multiple reasoning branches simultaneously.

Google DeepMind paper showing test-time compute scaling can outperform model scaling for difficult reasoning tasks. Demonstrates techniques for using compute at inference time (search, verification, revision) to dramatically improve model performance without training.

Paper introducing a two-stage prompting strategy: first plan the solution steps, then solve each step. Outperforms chain-of-thought on arithmetic and commonsense reasoning benchmarks without few-shot examples. The planning-first approach has become standard in modern agent orchestration.

Google Brain foundational paper showing that prompting LLMs to generate intermediate reasoning steps dramatically improves performance on complex tasks. The ancestor of all modern agent reasoning techniques. Introduced few-shot chain-of-thought and zero-shot "think step by step" prompting.

Microsoft Research paper introducing the AutoGen framework for multi-agent conversations. Demonstrates flexible conversation patterns where agents can be LLM-powered, human, or tool executors. Shows how multi-agent setups outperform single-agent approaches on complex coding and reasoning tasks.