Skip to content
Search

Intelligent agent

lastmod 2026-11-03created 2025-01-0182 edits

An agent is a system that tries to achieve goals in a complex, dynamic environment. An agent interacts with the external environment using sensors and actuators. A computer program that lives in a virtual environment and automates repetitive tasks to help users working with a particular application is called a software agent or an intelligent agent.

On the Internet, what is called an intelligent agent (or "agent") is a program that periodically gathers information or performs some other service without user intervention.

Categories

  • Agent Client Protocol (ACP)
  • Agentic Patterns - a collection of practical design patterns for AI agents
  • Manus - a general-purpose AI agent that connects thought and action
  • Opsy - an AI-powered SRE teammate
  • Korean: 스몰빌 (Smallville) - a virtual village where AI agents live
  • Lightpanda - a headless browser for AI and automation
  • Composio - an integration framework for AI agents
  • Microsoft AutoGen - a framework for next-generation LLM applications
  • BeeAI Framework - a framework for multi-agent development
  • Cua - open-source Docker containers for AI agents
  • Sim Studio - an open-source platform for building agent workflows
  • Oxy - an open-source framework for building SQL agents and data automation
  • Darwin Gödel Machine - an AI that evolves by modifying its own code
  • Agno - a high-performance framework for building multi-agent systems
  • Nxtscape - an open-source agentic browser
  • OpenCode - an AI coding agent for the terminal
  • Crush - a glamorous AI coding agent for the terminal
  • AGENTS.md - an open format for AI coding agents
  • Shadow - a background coding agent with a real-time web interface
  • Beads - a memory upgrade for coding agents
  • Dexto - an all-in-one toolkit for building agentic applications
  • Inkeep Agents - a platform for building AI agents with a No-Code builder and a TypeScript SDK
  • Mantic - a structural code search engine for AI agents
  • Clawdbot - a personal AI assistant that works on every OS and platform
  • Mastra - an AI framework from the Gatsby team
  • Plannotator - a tool for visually viewing and reviewing coding agents' plans
  • Entire - a new developer platform for AI agents
  • Cloudrouter - a skill that lets Claude Code/Codex spin up cloud VMs and GPUs
  • AionUi - an open-source multi-AI-agent desktop
  • agent-of-empires - a terminal session manager for AI coding agents using tmux and git Worktree
  • ChatDev - a Zero-Code development platform based on multi-agent collaboration
  • Agent Safehouse - a local agent sandboxing tool for macOS
  • page-agent - add an AI agent to a web page with one line of code
  • ralph-loop - automating development work with a long-running AI agent loop
  • onecli - a secret gateway for AI agents
  • OpenSquirrel - a GPU-based control panel for AI code agents
  • cq - Stack Overflow for agents
  • understudy - a local desktop agent that learns from demonstrations
  • MiroFish - a multi-agent-based prediction simulation engine
  • Optio - a workflow orchestrator for AI coding agents
  • Claw-Empire - an AI agent empire commanded from the CEO's desk
  • Paper Clip - open source that runs a company with AI
  • InsForge - an agentic backend platform for AI coding agents
  • goose - an open-source, locally run autonomous AI development agent from Block
  • HyperAgents - when agents design their own harness (Meta)
  • OpenHarness - an open agent harness with the built-in personal agent Ohmo
  • oh-my-agent - a practical multi-AI IDE agent harness
  • Mozilla Thunderbolt - an open-source AI client fully controlled by the user
  • CodeBurn - a TUI dashboard tracking token usage/cost of AI coding tools
  • Mindwalk - a visualization tool that replays coding agent sessions on a 3D map of the codebase
  • Hermes Agent
  • OpenQodex - a code review tool in which a separate agent verifies a coding agent's changes
  • Herdr

Other

AI agents - areas of opportunity

Various startups are developing innovative products using AI agents

Application examples

  • Healthcare: OpenClinic – a system that supports doctors
  • Robot education: innate – a robot training platform
  • Personal assistant: Khoj – acts as the user's second brain
  • Interior design: Rastro – acts as a personalized interior designer
  • Success story: HappyRobot
    • Automates phone calls and communication at logistics companies
    • Has 50+ customers already seeing tangible benefits. Average call time cut by 50%, operating costs reduced by 1/3

Tooling examples

  • No-code agent building: Gumloop – build AI agents without writing code
  • Phone payment system: Protegee – supports secure phone payments through agents
  • Security testing: uses "red team" AI agents to automatically stress-test a system's security and alignment

AI agent market map

Groups 138 AI agent companies broadly into Agent Orchestration and Application, and places them in subcategories

Development platforms

  • No/Low-Code platforms
    • AutoComputer, Autotab, Basepilot, Brainbase, Epsilla, Extend, Gumloop, Manaflow, Paradigm, Skyvern, Sola, Strada, Tasker ai
  • Code-First platforms
    • SciPhi
  • Testing & Evaluation
    • Blast, Coval, Fix AI, fixa, Foundry, Hamming AI, Laminar, MAIHEM, Relari, Synth, Vectorview

Runtime Services

  • Authentication & Security
    • Keet, HumanLayer, Abundant
  • Payments & Financial
    • Elevate, Protegee
  • Integration & Connectivity
    • Kura AI, reworks, Scrapybara, SID

Data & Operations

  • Memory & Knowledge Management
    • Zep AI
  • Governance & Compliance
    • FiddleCube
  • Observability & Monitoring
    • Vocera

Business Functions

  • Customer Support & Success
    • AgentsForce, Andoria, Cekura, RowBoat Labs, Marr Labs, Struct, Retell AI, Phonely, Markprompt, Duckie, Giga ML, Parahelp, PathPilot
  • Finance & Accounting
    • Concourse, Dimely, Fazeshift, Lighthouz, TaxGPT
  • Legal & Compliance
    • Greenlite, Sei, Sphinx
  • Research, Data & Analytics
    • AnswerGrid, Brewit, camelAI, MinusX, Reprompt, Reworkd, Simplifine, Undermind
  • Sales & Marketing
    • Apten, Coldreach, Fiber AI, Inventive AI, Kiosk, Mica AI, OpenFunnel, Origami Agents, Persana AI, Rankai, raycaster, revi, Seals AI, Summed AI, telli, Topo
  • Software Development
    • Codebuff, CodeStory, CodeViz, Decipher AI, Ellipsis, Gecko Security, MagiCode, OneGrep, Random Labs, Reflex, Speck, SRE.ai, Srcbook, Storia AI, Tempo Labs, vespper

Industry Solutions

  • Education & Training
    • Khoj
  • Financial Services
    • Arva AI, Cascading AI, CollectWise, Decisional AI, Finic, Kastle, Powder, Rulebase
  • Healthcare & Life Sciences
    • Astrix Health, Helpcare AI, Kairo Health, OpenClinic, Sample, Scritch, Shasta Health, ShowAndTell, Simbie AI, Substrate
  • Industrial & Robotics
    • Dime, Entangl, Forerunner AI, Innate, PraxisAI, Tiny
  • Legal Services
    • Pearson Labs
  • Leisure & Hospitality
    • Gander
  • Logistics & Supply Chain
    • Dexter, HappyRobot, Linc., Mandel AI, Pulse AI
  • Media & Entertainment
    • Ego
  • Retail & E-commerce
    • AI Sell, Anglera, pap!, PurplePages, Rastro, Toma, Yuma AI

Open-source AI agent frameworks

An article comparing various open-source AI agent frameworks such as LangGraph, OpenAI Agents SDK, Smolagents, CrewAI, AutoGen, Semantic Kernel, and LlamaIndex Agents.

  • AI agent development used to be a combination of scripts, prompt engineering, and trial and error, but open-source frameworks are now growing in number
  • Each framework has its own architectural philosophy balancing agent autonomy and stability
  • Integrating with observability tools like Langfuse lets you visualize and debug prompts, responses, and API calls

Solution comparison

  • LangGraph – graph-based workflows
    • A graph-architecture framework built on LangChain
    • Each node handles a prompt or unit of work, and edges control data flow and branching
    • Well suited to complex multi-step tasks, parallel processing, and inserting error handling logic
    • Strong at visualization and debugging; suitable for state-based agent design
  • OpenAI Agents SDK – the official OpenAI agent toolkit
    • Official SDK provided by OpenAI
    • Integrates naturally with models such as GPT-4o and GPT-o3
    • Can perform multi-step tasks by configuring roles, tools, and triggers
    • Suited to users familiar with the OpenAI ecosystem
  • Smolagents – a minimal, code-based approach
    • Hugging Face's minimal, code-centric agent framework
    • The AI generates and runs Python code in a simple loop
    • Suited to rapid prototyping without complex orchestration
    • Uses ReAct-style prompts internally
  • CrewAI – role-based multi-agent collaboration
    • Assigns each agent a distinct role so they can collaborate
    • Automatically coordinates workflows through a container concept called a "Crew"
    • Makes it easy to implement scenarios such as Planner - Researcher - Writer
    • Includes memory features and error handling logic
  • AutoGen – asynchronous conversational agents
    • An asynchronous, conversation-based agent framework developed by Microsoft Research
    • Each agent operates asynchronously by exchanging conversation messages
    • Well suited where multi-turn conversations, role switching, and real-time tool calls matter
    • Its event-driven structure suits high-concurrency tasks
  • Semantic Kernel – enterprise-friendly workflows
    • Microsoft's .NET-centric framework
    • Builds plan-based workflows by mixing AI skills and code-based skills
    • Optimized for enterprise requirements such as security, compliance, and Azure integration
    • Supports multiple languages including Python, C#, and Java
  • LlamaIndex Agents – data-centric agents
    • LlamaIndex started as a RAG-based framework and extended into agent features
    • Searches local/external data sources and links the results to responses or actions
    • Suited to document-based Q&A, summarization, and custom search agents
    • Low barrier to entry if you have used LlamaIndex before

Which framework to use when?

  • Task complexity: choose a suitable framework depending on whether the task is simple or complex
    • Complex multi-step flows: LangGraph, Semantic Kernel
    • Simple code-based execution: Smolagents
  • Multi-agent collaboration: when multiple agents are needed, you need an architecture that supports asynchronous conversation and role delegation
    • Role-based agents: CrewAI
    • Conversational asynchronous agents: AutoGen
  • Integration environment: consider the environments and systems the agent must interact with
    • OpenAI-centric services: OpenAI Agents SDK
    • Need to integrate with existing business logic: Semantic Kernel
  • Performance and scalability: consider the application's performance requirements. If real-time interaction is needed, an event-driven architecture may be required
    • Need for high-concurrency processing: AutoGen
    • Integrate with observability tools (e.g. Langfuse) for debugging and tracing

Beyond the AI MVP: what it really takes

The AI MVP trap most teams fall into:

  1. Stage 1. The deceptive MVP: a demo that looks finished in just a few days. But it soon starts making basic mistakes
  2. Stage 2. ±0: you try to improve it, but it only gets more complex and becomes unpredictable. Most companies (90%) are at this stage
  3. Stage 3. Science: at this point you realize you need evaluation tests, observability tools, etc., and start building them
  4. Stage 4. It actually works: you finally have continuous monitoring, a comprehensive evaluation test suite, and fast analysis tools

OpenAI's practical guide to building agents

See OpenAI#OpenAI's practical guide to building agents.

Emerging developer patterns for the AI era

A development culture is forming that treats AI not as a mere tool but as a foundational technology.

  • Existing approaches to version control, dashboards, templates, documentation, secret management, etc. are changing for the AI era
    • Git is being reinterpreted as state management centered on prompts and test results rather than line-by-line code changes
    • Agents become both authors and consumers of code, increasing the need to redesign the tools themselves
    • Dashboards and UIs are shifting to natural-language interfaces, evolving into something used by both humans and AI agents
    • Documentation is turning into a knowledge base for both humans and AI, restructured into formats agents can understand

The following nine patterns show that this is not simply layering AI on top of existing development practices, but that the way software is built is itself being redefined around agents, context, and intent

  1. AI-Native Git: redefining version control for AI agents
  2. Dashboards → Synthesis: evolving into dynamic AI-driven interfaces
  3. Docs are evolving into a combination of tools, indexes, and interactive knowledge bases
  4. From templates to generation: vibe coding replacing create-react-app
  5. Beyond .env: secret management in agent-centric environments
  6. Accessibility as the universal interface: apps through the eyes of an LLM
  7. The rise of asynchronous agent work
  8. MCP: Model Context Protocol, one step closer to a universal standard
  9. Abstracted primitives: the authentication, payments, and storage every AI agent needs

We put a coding agent in a while loop and it shipped 6 repositories overnight

Ran Claude Code headless in an infinite loop, repeatedly.

  • With this approach, they succeeded at various ports, such as automatically converting the assistant-ui React project to Vue and a Python project to TypeScript
  • Found that keeping prompts simple improves performance, while making them complex increases inefficiency
  • Though not perfect, they also built RepoMirror, a tool useful for syncing source/target repositories
  • They also observed unexpected behaviors and learning by the AI agents, such as stopping themselves and adding tasks, giving a firsthand sense of both the potential and limits of future automation

Running agents in an infinite loop

  • A form recommended by Geoff Huntley, in which a coding agent prompt is run continuously from the shell
    • Example: while :; do cat prompt.md | claude -p --dangerously-skip-permissions; done
  • Applied this loop approach to tasks such as converting assistant-ui from React to Vue
  • Limits of full automation
    • Notable problem: often produces ported code that does not fully work (e.g., some browser demos incomplete)
    • Prompt calibration and interactive fixes needed
  • Infrastructure/cost and operations - total cost about $800, 1100 commits in total, around $10.50/hour per agent
  • They experienced firsthand that simplicity is effective

See also

Favorite site

Guides

References

  1. :File:AIAgentsMarketMap.png - the thumbnail is broken, so view only the original file ↩

author Zer0 sitemap