Subject
106 entries
AI Agents
Bookmarks
OpenClaw Arena: real-world AI agent benchmarking
Public benchmarking platform that pits AI agents against each other on real tasks. Measures both performance and cost-effectiveness.
Claw-Code: cleanroom agent harness reimplementation in Rust
Cleanroom Rust reimplementation of an AI agent harness, built from reverse-engineered architecture. Hit 50K GitHub stars in 2 hours.
Claude Code Unpacked: reverse-engineering the agent harness
Deep reverse-engineering of Claude Code's internals — the agent loop, 50+ tools, and multi-agent orchestration. The closest thing to a technical spec for how it actually works.
Baton: parallel AI agent orchestration via git worktrees
Desktop app for running multiple AI coding agents in parallel using git worktrees. Solves the coordination problem that kills most multi-agent dev workflows.
AgentFlow: dependency graph orchestration for AI agents
Orchestrate thousands of AI agents as dependency graphs with parallel fanout and remote execution. Like make/Airflow but for LLM agents.
Agent Super Spy: LLM observability for local agent development
Local LLM proxy stack that gives you full visibility into what your AI agents actually do. Dual-layer observability via OpenTelemetry traces and raw HTTP capture.
rses: cross-agent session handoff
rses lets you transfer a full session context — task, git diff, conversation history — between Claude Code, Codex CLI, and OpenCode. Useful when you want to pick up where one agent left off in a completely different tool.
Cog: cognitive architecture for Claude Code
Cog is a cognitive architecture framework for Claude Code from lab.puga.com.br. Limited public documentation available — the name suggests structured reasoning or memory layers on top of the base Claude Code harness.
Claude Code Auto Dream: automated memory consolidation
Auto Dream is Anthropic's memory consolidation feature for Claude Code — a four-phase process that runs after sessions to merge, deduplicate, and prune memory files. Think of it as automated housekeeping that keeps your AI's long-term memory from accumulating contradictions.
Remind: generalized memory consolidation for AI agents
Remind consolidates raw AI experiences into generalized concepts stored in a knowledge graph, rather than dumping everything into a vector database. The distinction matters: instead of retrieving raw transcripts, it surfaces patterns like 'user prefers statically typed languages.'
Fabro: AI agent workflow graph orchestrator
Fabro defines AI agent workflows as Graphviz DOT graphs with git checkpointing, cloud sandboxes, and human-in-the-loop gates. It occupies the space between micromanaging an agent line-by-line and blindly accepting a 500-line diff.
NVIDIA Agent Toolkit: enterprise AI agent platform
NVIDIA launched an open-source Agent Toolkit with a runtime, enterprise data blueprint, and open models that cut query costs by 50%+. The company frames this as an 'industrial revolution in knowledge work' — which is either accurate or marketing, depending on how the next few years go.
Ink: infrastructure designed for AI agents
Ink gives AI agents a cloud deployment skill — deploy services, configure domains, manage infrastructure autonomously. One API call to give an agent the ability to ship code to production.
GitAgent: git-native open standard for AI agents
GitAgent defines AI agents as version-controlled files in a git repo — SOUL.md, SKILL.md, agent.yaml — exportable to any framework. The bet is that agent definitions should live in your repo, not a vendor's cloud.
TinyFish: enterprise infrastructure for AI web agents
TinyFish is serverless infrastructure for AI web agents — single API, parallel execution at scale, 98.7% success rate, $0.04/op. Specifically built to handle authenticated websites and dynamic content that breaks standard scraping tools.
Axe: lightweight CLI for single-purpose AI agents
Axe is a minimal CLI for single-purpose LLM agents defined in TOML — composable like Unix pipes, triggerable from git hooks or cron, supporting Anthropic/OpenAI/Ollama. The anti-chatbot: each agent does one thing.
Kernel Managed Auth: secure authentication for AI agents
Kernel's Managed Auth gives AI agents a secure way to handle login flows — 2FA, SSO, 1Password integration — without ever exposing credentials to the LLM. Solves one of the least-discussed blockers for autonomous web agents.
Codex Infinity: hosted autonomous coding agent platform
Codex Infinity is a cloud platform for running autonomous coding agent sessions with advanced reasoning, AI art generation, and cloud orchestration. A hosted take on what local coding agents do, at unlimited scale.
Arize AX CLI: LLM observability for AI coding agents
Arize AX CLI makes LLM traces, spans, and experiments machine-readable so AI coding agents can run the same observability analysis as the browser UI. Closes the loop between observability data and the agents that produced it.
zclaw: ESP32-resident AI assistant
zclaw is an AI assistant that runs on an ESP32 microcontroller within a strict 888 KiB firmware budget. Handles natural language scheduling, GPIO control, memory, and tool calling — written in C with ESP-IDF.
Polyscope: cockpit interface for AI-driven development
Polyscope positions itself as 'the new cockpit' for AI-driven development — a central control interface for the era when AI writes code and humans supervise. Minimal public detail, but the framing is noteworthy.
OpenClaw Exposure Watchboard: tracking exposed AI agent deployments
The OpenClaw Exposure Watchboard tracks 569K+ publicly reachable AI agent instances — many with leaked credentials, CVEs, and active threat actor interest. Useful for checking if your own deployment is inadvertently public.
Nominal: AI agents for autonomous production incident response
Nominal is AI agents that read your code, watch your infra, and fix production before you wake up — eliminating on-call by 2026. Early access, integrates with AWS/Cloudflare/Vercel/Fly.io and more.
The Self-Driving Codebase: background agents for enterprise delivery
The Self-Driving Codebase thesis: individual developer speed from laptop agents doesn't translate to org velocity. Background agents need cloud isolation, event routing, and governance to actually move DORA metrics.
cmux: native macOS terminal for parallel AI coding agents
cmux is a native macOS terminal built on Ghostty specifically for running multiple AI coding agents in parallel. Sidebar tabs show which agents need attention, with notification rings and desktop alerts when an agent is waiting for input.
Agent Swarm: persistent multi-agent coding framework
Agent Swarm is a lead/worker AI coding agent framework with persistent identity files, session memory extraction, and Docker isolation. Each agent accumulates expertise across sessions through SOUL.md, IDENTITY.md, TOOLS.md, CLAUDE.md.
Tessl: package manager for AI agent skills
Tessl is npm for AI agent skills — versioned, evaluated, and discoverable through a public registry. Skills can be submitted via GitHub URL or evaluated locally via CLI before publishing.
Tessl content-strategy skill: community agent skills in practice
A Tessl-registered content-strategy skill by coreyhaines31 with 81% quality rating — an example of community-contributed, evaluated agent skills in the Tessl registry. Shows what agent skill distribution looks like in practice.
Paper: connected canvas for design-to-code with AI agents
Paper is a design tool that exports directly to HTML/CSS code and connects to AI agents (Claude, Copilot) for real-time collaboration. Positions itself as closing the design-to-development gap through agent-assisted workflows.
Shuru: local-first microVM sandbox for AI agents
Shuru is a local-first microVM sandbox for AI agents on macOS — gives agents an isolated environment to execute code without touching the host system. Addresses the core risk of agents with shell access running on your actual machine.
Exploitation-Validator: LLM sub-agent vulnerability analysis
exploitation-validator uses LLM sub-agents to find and validate code vulnerabilities through a five-stage pipeline that explicitly cross-checks findings against actual source code to eliminate hallucinations before generating PoC exploits.
Cord: dynamic task trees for coordinating AI agents
Cord is a framework for dynamic task trees of AI agents — agents decompose goals into subtasks at runtime, not at design time. Built on five MCP primitives with spawn/fork context primitives for controlling information flow between agents.
Mnemom: proving what AI agents are thinking
Mnemom.ai makes AI agent reasoning transparent — "Prove What Your AI Agents Are Thinking." An agent explainability tool for validating agent decision-making.
dmux: parallel AI agents in tmux worktrees
dmux lets you run multiple AI coding agents in isolated git worktrees, managing branch/develop/merge cycles in parallel. Another take on the same problem as Baton and cmux, specifically oriented around tmux-based workflows.
visual-explainer: HTML pages instead of ASCII art from agents
visual-explainer is a Claude Code agent skill that generates styled HTML pages instead of ASCII art for diagrams, diff reviews, architecture overviews, and data tables. Fixes the terminal readability problem for complex agent output.
Nightshift: overnight codebase maintenance with leftover token budget
Nightshift uses your leftover Claude/Codex token budget to run overnight code maintenance — dead code, doc drift, test gaps, security issues — landing everything as PRs. Like a Roomba for your codebase.
WebMCP Starter: demo for the WebMCP browser standard
webmcp-starter is a demo app for WebMCP — a proposed web standard that exposes structured tools for AI browser agents directly from web pages. A food delivery app with 9 WebMCP tools demonstrates the full ordering flow.
Safe Yolo Mode: running LLM agents in isolated VMs
A guide to running LLM agents with full auto-approval permissions inside libvirt/virsh VMs, so risky operations are contained without restricting what the agent can do. The sandbox is the safety rail, not permission restrictions.
Kimi Claw: 24/7 AI assistant with long-term memory
Kimi Claw is Moonshot AI's 24/7 AI assistant with long-term memory and automation — deep research, document processing, coding, and cross-platform availability. 4.9 stars across 94,000+ reviews.
Zo: personal AI cloud computer for 24/7 execution
Zo is a personal AI cloud computer that runs autonomously 24/7 — manages email, books meetings, builds apps, handles files — via natural language text instructions. A personal cloud computer with full file access and multi-model AI.
VS Code PR: LSP linting for AI prompt files
VS Code PR that adds LSP-based real-time linting for AI prompt files — flags weak instructions, ambiguity, XML structure issues, redundancy, template variable errors, and token budget problems. Like ESLint for prompts.
Letta Context Repositories: git-versioned memory for coding agents
Letta introduces Context Repositories — git-versioned memory for coding agents that treats context as source code with branches, commits, and diffs. A substantive rethink of how stateful coding agents should manage long-lived context.
lean-collab: multi-agent theorem proving in Lean 4
lean-collab applies multi-agent collaboration to Lean 4 theorem proving, using the Ensue Memory Network to coordinate agents across complex mathematical proofs. An unusual intersection of formal verification and multi-agent AI.
IronClaw: secure personal AI assistant from NEAR AI
IronClaw is a privacy-focused personal AI assistant from NEAR AI, implemented in Rust with WASM sandboxing for untrusted tools, credential protection at the host boundary, and hybrid search memory. Your data never leaves your control.
Hive: production runtime harness for AI agents
Hive (by Aden, YC-backed) is a production runtime harness for AI agents: state management, checkpoint-based crash recovery, cost enforcement, and self-healing through graph evolution. 102 MCP tools.
roro: workspace manager for parallel AI coding agents
roro is a workspace manager for multiple parallel AI coding agents — Agent Calendar for scheduling overnight work, real-time activity monitoring via Claude Code hooks, and personalized agent identities per workspace.
Entire: $60M seed round, Checkpoints product
Entire launched with a $60M seed round and their first product, Checkpoints — a tool for AI-assisted software development. The seed round size signals serious investor conviction in agent-driven development infrastructure.
The Companion: web UI for Claude Code and Codex
The Companion is a web UI for Claude Code and Codex sessions — parallel agents, full tool call timeline, permission gating, and session recovery. Replaces terminal tab management for multi-agent workflows.
tapes: transparent AI agent telemetry
Tapes introduces transparent telemetry for AI agents — a structured trace format that makes agent execution auditable without requiring a full observability platform. Good complement to heavier tools like Arize AX or Phoenix.
Searchable Agent Memory in a Single File
A single-file BM25 MCP server that gives a Claude Code agent searchable access to its own conversation history. Shows how much mileage you can extract from a minimal implementation.
Mastra Observational Memory
Mastra's Observational Memory keeps agent context windows small by summarizing and compressing observations rather than keeping raw conversation history. A practical approach to the long-term vs. working memory tradeoff in AI agents.
Destructive Command Guard (dcg): hook for blocking dangerous agent commands
Destructive Command Guard (dcg) is a sub-millisecond hook that intercepts dangerous commands before AI agents execute them — git resets, rm -rf, DROP TABLE, kubectl deletes. 49+ security packs, SIMD-accelerated, never false-blocks.
Configurancy: keeping systems intelligible when agents write all the code
Kyle Mathews argues that AI makes code cheap, so the scarce asset is now system self-knowledge: explicit contracts, conformance suites, specs. Configurancy is the term for what lets bounded agents coherently co-evolve a shared system.
Matchlock: AI agents in ephemeral microVMs
Matchlock runs AI agents in ephemeral microVMs with network allowlisting and MITM-based secret injection — your API keys never enter the VM. Full Linux environment that boots under a second, vanishes when done.
Plano: AI-native proxy and data plane for agents
Plano is an AI-native proxy and data plane for agentic apps — centralized routing, guardrails, observability, and LLM switching so you don't re-implement plumbing in every codebase. Decouples agents from brittle framework abstractions.
roborev: continuous code review for AI agent commits
roborev automatically reviews every commit an AI coding agent makes via git hooks, accumulating findings in a persistent queue. The problem it solves: agents write fast but move on before reviewing their own mistakes.
Pi Messenger: file-based multi-agent coordination for Pi
Pi Messenger turns multiple AI coding agents in different terminals into a chat room — agents can see who's online, claim tasks, reserve files, and send messages. Coordination via plain files, no daemon required.
Augment Intent: coordinator-plus-specialist agent orchestration
Augment Intent is a desktop app where a coordinator agent breaks tasks into living specs, then delegates to specialist agents working in parallel. The specs stay current as code evolves — avoiding the documentation decay that makes most spec-driven workflows impractical.
Agentboard: web-based tmux GUI for AI agent sessions
Agentboard is a web GUI for tmux optimized for AI agent TUIs — run Claude Code or Codex on your server and interact from your phone via iOS Safari. Status inference, session pinning, and a shared workspace across devices.
vmux: sandbox runtime for AI agents
vmux is a sandbox runtime for AI agents — preview URLs, log tailing, and tmux attach, running on Cloudflare CPU with Modal GPU backends. Gives agents a managed execution environment rather than direct host access.
Honcho: continual learning memory for AI agents
Honcho is an AI-native memory platform that gives agents continual, reasoned learning rather than raw retrieval — 60-90% token savings by surfacing only essential context. Background 'Dreaming' keeps understanding current without runtime overhead.
Skyll: runtime skill discovery for AI agents
Skyll is a REST API and MCP server that lets any AI agent discover and learn skills at runtime, without pre-installation. Aggregates SKILL.md files from GitHub and returns structured JSON for context injection.
Dash: self-learning data agent with 6-layer context
Dash is a self-learning data agent from Agno that grounds answers in 6 layers of context, inspired by OpenAI's in-house implementation. Each query improves the agent's performance via adaptive memory.
AgentSkills.io: open format for agent capabilities
AgentSkills.io is a simple open format for giving AI agents new capabilities and expertise — a registry and spec for agent skill files. Makes skills portable and discoverable across agent frameworks.
Zenflow: spec-driven orchestration for AI development
Zenflow is Zencoder's orchestration platform for AI development — spec-driven workflows, automated verification, and multi-agent coordination. Tries to make AI-generated code reliable at scale.
sandbox-agent: universal sandbox API for coding agents
sandbox-agent is a universal API for running coding agents in sandboxes — works with Claude Code, Codex, OpenCode, and Amp with a single interface. Rivet's contribution to the fragmented coding agent sandbox space.
Executive: human-centered AI agent control
Executive is a GitHub app for maximizing developer control over AI agents — human executive control, not agent autonomy. Developer directs, agents execute, and the human stays in strategic command.
EchoVault: local-first persistent memory for coding agents
EchoVault is local-first persistent memory for coding agents — decisions, bugs, and context stored as Markdown files indexed with FTS5, served via MCP. No cloud, no RAM overhead at idle, no external servers.
Aviator Runbooks: multiplayer AI coding via shared runbooks
Aviator Runbooks makes AI coding multiplayer — reusable, shareable AI runbooks for teams that can automate PR reviews, standardize workflows, and integrate with existing AI tools. Specs become team-shared playbooks.
Supacode: terminal command center for parallel coding agents
Supacode is a native terminal command center for running 50+ AI coding agents in parallel. Designed for developers who want to run fleets of agents from a single terminal interface.
HyperContext: spatial context visualization for AI sessions
HyperContext visualizes AI session state as a spatial map — see what the model actually knows, how the context window is structured, and where attention is distributed. Makes the invisible context window visible.
amla-sandbox: WASM capability sandbox for AI agents
amla-sandbox is a WASM sandbox with capability enforcement for AI coding agents — no Docker, no VM, one binary. Agents can only call explicitly provided tools. Addresses the arbitrary code execution risk in most agent frameworks.
AgentRC: context engineering and AI-readiness for repos
AgentRC (formerly Primer) measures and generates AI-readiness for repos — scoring across 9 pillars, generating instruction files, and running CI drift detection so agent context stays current as code evolves.
Open-Inspect: open-source background coding agent system
Open-Inspect is an open-source background coding agent system inspired by Ramp's internal Inspect tool — agents work in full dev environments while you're elsewhere, then create PRs. Supports multiplayer sessions and Anthropic or OpenAI models.
Retain: unified AI conversation knowledge base for macOS
Retain is a native macOS app that aggregates AI conversations from Claude Code, Cursor, ChatGPT, and others into a unified searchable knowledge base, then exports learnings to CLAUDE.md so Claude remembers your preferences.
Descope: identity platform for humans and AI agents
Descope is an External IAM platform that handles identity for both human users and AI agents — the same system covers user auth, business customer identity, MCP server auth, and agentic workflows. Identity as the control plane for human+agent systems.
Documentor: documentation aggregator for AI assistants
Documentor is a tool for aggregating external documentation into a format optimized for LLM consumption — crawling docs sites, building search indexes, and storing structured outputs that AI assistants can query during development work.
Dexto: intelligence layer for AI agents
Dexto is an open-source AI agent orchestration platform that adds stateful memory, tool management, and error recovery to LLMs — supporting 50+ models, 30+ MCP tools, and deployment across web UI, CLI, REST API, and Discord. A production-oriented agent infrastructure layer.
Advanced Context Engineering for Coding Agents
A practical guide to making AI coding agents effective in large production codebases through deliberate context management — research/plan/implement phases, 40-60% context utilization, and subagents for isolation. The author shipped 35k LOC in 7 hours using these techniques.
I Built 6 AI Employees in 10 Minutes
A YouTube video demonstrating building 6 autonomous AI agents ('employees') in 10 minutes that run 24/7 without human intervention. Likely covers no-code or low-code agent orchestration tools and common automation patterns.
Par: Parallel Worktree and Session Manager
Par is a CLI that combines git worktrees with tmux sessions to manage parallel development workstreams — each task gets an isolated directory and persistent terminal. Built for concurrent AI-assisted development where multiple agents run simultaneously.
CCPM: Spec-Driven Project Management for Claude Code
CCPM is an agent skill that orchestrates spec-driven software development using GitHub Issues as project state and git worktrees for parallel agent execution, reducing wall time 5x by running independent tasks simultaneously. The anti-vibe-coding project manager.
Leap: AI Developer Agent That Builds and Deploys to Your Cloud
Leap is a full-stack AI developer agent that builds and deploys applications directly to your cloud (AWS or GCP) — you describe what to build, it codes and deploys. It's the most integrated version of the AI coding agent vision: not just code generation but end-to-end deployment.
Clutch Security: Non-Human Identity Security
Clutch Security specializes in Non-Human Identity (NHI) security — protecting service accounts, API keys, machine identities, and now AI agents across digital environments. NHI is becoming the dominant attack surface as automated systems proliferate.
Terragon Labs: Background Agents for Claude Code
Terragon Labs runs Claude Code agents in parallel inside remote sandboxes, letting you automate multiple development tasks concurrently without managing infrastructure. It's a managed background-agent platform for Claude Code.
How to Build an Agent: Minimal Code-Editing Agent in <400 Lines
A practical tutorial showing how to build a fully functional code-editing agent in under 400 lines of code. Worth reading as a minimal reference implementation that strips the problem down to its essential loop.
Sim Studio: Visual Agent Workflow Builder
Sim Studio is an open-source visual agent workflow builder — connect LLMs, tools, and vector stores on a canvas, then deploy immediately. An AI copilot generates nodes and fixes errors from natural language, and 1000+ integrations are supported.
Advanced AI Agents Course (DAIR.AI)
DAIR.AI's Advanced AI Agents course covers sophisticated agentic patterns — prompt chaining, routing, parallelization, multi-agent architectures, evaluator-optimizer patterns, and deployment on Google Cloud Run. 38 lessons, 4.5 hours, taught by Elvis Saravia.
Keycard: Identity and Access Control for AI Agents
Keycard is a security control plane for AI agents — every action an agent attempts is evaluated against policy before credentials are issued. Short-lived scoped tokens replace long-lived permissions, and every agent interaction is auditable back to a human identity.
Autonomous Hacker: LLM-Powered Security Research Agent
The autonomous-hacker module in R3DRUN3's sploitcraft demonstrates using LLMs as autonomous security research agents — scanning, identifying vulnerabilities, and generating exploits with minimal human input. An educational proof of concept for LLM-assisted offensive security.
ai-llm-agent-solver: Autonomous Gandalf Challenge Solver
An LLM-powered agent that autonomously solves the Gandalf AI challenge — a prompt injection security game where you try to extract a secret password from a guarded AI. Uses OpenAI API and agent-based reasoning.
DSPy Self-Discover Framework
A DSPy + Groq-powered framework that takes natural language task descriptions and generates executable Python code to solve them — automated problem-solving through AI-driven reasoning and code generation. Now archived.
TravelPlanner: A Benchmark for Real-World Planning with Language Agents
TravelPlanner is a benchmark for evaluating LLM planning capabilities in complex real-world scenarios — GPT-4 scored 0.6% on the full benchmark, revealing that even the best LLMs struggle with multi-constraint sequential planning. A sobering check on agentic AI ambitions.
Eidolon: open-source agent service framework
Eidolon is an open-source agent service framework that treats AI agents as first-class services with defined APIs, not just Python scripts — enabling agents to communicate with each other and be composed into larger systems. Targets production deployment of multi-agent architectures.
Memoria: structured memory management for AI agents
Memoria is a memory manager for AI agents that routes information to appropriate memory types (episodic, semantic, procedural) to enable dynamic, context-aware responses. An attempt to implement a structured cognitive architecture for LLM memory.
Diff History for Neural Language Agents
Diff history applies the Unix diff command to sequential agent observations, replacing full text states with change deltas. Dramatically reduces context length while preserving critical information — small models tuned with diff history matched SOTA on NetHack with 1800x fewer training examples.
TokenCost: LLM API Cost Calculator
TokenCost is a Python library from AgentOps that counts tokens and calculates USD costs for 400+ LLM models before making API calls. Keeps a live-updated pricing database so your cost estimates don't go stale when providers update pricing.
Internet of Agents: Browser Agent Swarms via MultiOn
Internet of Agents is an open-source framework for building swarms of internet-browsing AI agents using MultiOn, enabling autonomous web tasks at scale. An early demonstration of multi-agent browser automation coordinated through a single API.
The Rise and Potential of LLM-Based Agents: A Survey
An 86-page survey paper by Zhiheng Xi et al. comprehensively mapping the architecture, capabilities, and applications of LLM-based agents — establishing the planning/memory/tool-use framework that became the standard way to think about agent components. The field's foundational survey document.
LLM Agent Paper List
A curated paper list accompanying the 86-page 'Rise and Potential of LLM-Based Agents' survey — organized by agent component (planning, memory, tool use, action) and application domain. The most comprehensive catalog of LLM agent research from September 2023.
agentaction: Action Chaining and History for LLM Agents
agentaction is a Python library for action chaining and history management in LLM agents — a lightweight abstraction for defining, executing, and tracking sequences of agent actions with persistent history. An early building block for agent frameworks.
LLM Powered Autonomous Agents
Lilian Weng's survey post on LLM-powered autonomous agents — covering the planning, memory, and tool use components that compose into agent architectures. One of the most cited and comprehensive overviews of the agent design space from mid-2023.
JARVIS / HuggingGPT: LLM as AI Model Orchestrator
Microsoft JARVIS (also published as HuggingGPT) uses ChatGPT as a task planner that routes subtasks to specialized Hugging Face models — an early demonstration that LLMs could orchestrate other AI models as tools. A prototype of the multi-model agent pattern.
ChatArena: Multi-Agent Language Game Environments for LLMs
ChatArena provides multi-agent game environments for LLMs — structured settings where multiple LLMs interact, debate, negotiate, or play games with defined rules and roles. A research framework for studying emergent multi-agent behavior.
Jim Fan: Foundation Models for Embodied Agents
Jim Fan's talk on foundation models for embodied agents — covering how large pretrained models can be adapted for physical and simulated agents that act in the world. Previews the Voyager and MineDreamer research that followed from his NVIDIA lab.
NatBot: GPT-3-Driven Browser Automation
NatBot is an early LLM-driven browser automation bot by Nat Friedman — it takes a natural language task, drives a Chromium browser via Playwright, and completes multi-step web tasks. A 2022 proof-of-concept for what later became the AI browser agent category.