AI Agents Explained: What They Are and 15 Tools Compared (2026)
February 18, 2026 · Updated September 28, 2026 · 12 min read
An AI agent is software that is given a goal and works toward it by taking actions: opening web pages, running code, calling APIs, editing files, sending messages. A chatbot answers; an agent does. This guide explains the difference, sorts the current tools into four kinds, and lists what each one costs. Every product and price here was checked September 2026 against the vendor's own site, and the pages we used are listed at the end.
The mechanism is a loop. The agent looks at the current state (a screenshot, a page, a file, an API response), decides on an action, carries it out, checks what changed, and repeats until the goal is met or it hits something it should ask a person about. The model inside is usually the same one you would chat with. What makes it an agent is the scaffolding around the model: tools it is allowed to call, permissions on what it can touch, and a runtime that keeps the loop going.
Since this guide was first published the market has consolidated. OpenAI folded Operator into ChatGPT, Zapier renamed Central to Agents, Google spread its agent features across Gemini and Chrome, and coding agents became the category most people actually use every day. The 15 tools below are grouped into browser and computer-use agents, coding agents, workflow builders and developer frameworks. For AI tools beyond agents, browse our AI productivity tools category.
| Tool | Kind and use | Pricing |
|---|---|---|
| ChatGPT agent mode (OpenAI) | Browser agent: everyday web tasks inside ChatGPT | Paid ChatGPT plans, from $20 a month |
| Claude in Chrome and computer use (Anthropic) | Browser agent for paid Claude plans, plus a desktop-control API | Paid Claude plans from $20 a month; API usage for computer use |
| Gemini Spark and Chrome auto browse (Google) | Account agent: Gmail, Calendar and the web on a Google plan | Google AI Pro at $19.99 a month, or Ultra |
| Manus | General-purpose cloud agent that works in its own sandbox | Free tier; Pro from $20 a month |
| Claude Code (Anthropic) | Coding agent: terminal, IDE, desktop app and browser | Claude Pro from $20 a month, or API usage |
| Codex (OpenAI) | Coding agent included with every ChatGPT plan | All ChatGPT plans, including Free |
| Microsoft Copilot Studio | Workflow builder: governed agents for Microsoft 365 and Teams | $200 per 25,000 credits a month, or pay as you go |
| Zapier Agents | Workflow builder: agents across Zapier integrations (formerly Zapier Central) | Free tier; Agents Pro from $33.33 a month |
| Lindy | Workflow builder: Slack-first AI teammate for recurring office tasks | From $29.99 per user a month; no free plan |
| n8n | Workflow builder: open source, with agent nodes and free self-hosting | Free self-hosted; cloud from 20 euros a month |
| Relevance AI | Workflow builder: low-code specialist agents for sales and operations | From $19 a month; free plan retired |
| AutoGPT Platform | Workflow builder: the 2023 open-source agent, rebuilt as a visual platform | Free self-hosted; hosted from $42.50 a month |
| LangGraph | Framework: durable, stateful agents with human checkpoints | Open source; LangSmith from $0 per seat |
| CrewAI | Framework: multi-agent crews with roles and tools | Open source; hosted Basic plan free |
| OpenAI Agents SDK | Framework: minimal, provider-agnostic agent library | Open source; pay for model usage |
15 AI Agent Tools, Grouped by Kind
The list runs from tools you can use today with no setup (browser and account agents), through coding agents, to workflow builders for teams, and finally frameworks for engineers building their own. Prices are the vendor's list prices in US dollars unless stated, and most vendors meter usage in credits or requests on top of the subscription, so the plan name tells you less than the limits do.
1. ChatGPT agent mode (OpenAI)
Browser agent: everyday web tasks inside ChatGPT
Agent mode is the successor to Operator, which OpenAI folded into ChatGPT in 2025; the standalone Operator site no longer exists. You describe a goal and ChatGPT works in its own virtual browser: it opens sites, reads pages and fills forms, and it combines that with deep research and code execution before reporting back. It asks for confirmation before purchases, sending messages and similar actions, and you can take over the browser at any point. Each agent request counts against a monthly limit that varies by plan. It is the easiest browser agent to try if you already pay for ChatGPT, and it handles comparison shopping, form filling and research well. It struggles with sites that block automation or need unusual logins.
2. Claude in Chrome and computer use (Anthropic)
Browser agent for paid Claude plans, plus a desktop-control API
Anthropic offers agents at two levels. Claude in Chrome is a browser extension that reads the page you are signed in to, then clicks, types and fills forms while you decide what happens next; it is included on every paid Claude plan. For developers, the computer use tool in the Claude API gives the model screenshot, mouse and keyboard control of a desktop environment you host, so it can work across any application rather than only a browser. Computer use is an API feature: you run the virtual machine, your code executes each action, and the model reads the result and decides the next step in a loop. It is more flexible than a browser-only agent and correspondingly more work to set up.
3. Gemini Spark and Chrome auto browse (Google)
Account agent: Gmail, Calendar and the web on a Google plan
Google has folded agent features into its subscriptions rather than shipping one product. Gemini Spark is a personal agent inside the Gemini app that automates workflows and scheduled tasks across your Google account, such as summarising or archiving newsletters, unsubscribing from lists, or keeping a custom news digest up to date. It asks for confirmation before sending communications, changing data, making purchases or submitting web forms. Separately, auto browse in Gemini in Chrome completes multi-step tasks on the web, such as comparing products or booking travel, and is limited to 20 requests a day on Pro and 200 on Ultra. Both are gated by region: Spark is not offered in the EEA, UK, Switzerland or Nigeria, and auto browse is US only at the time of writing.
4. Manus
General-purpose cloud agent that works in its own sandbox
Manus is a hosted general-purpose agent. You give it a task, such as building a comparison spreadsheet, researching a market or producing a slide deck, and it works in a cloud sandbox with its own browser, file system and code execution, then hands back the result. Because it runs remotely you can close your laptop and check back later. Work is metered in credits, and a long task with a lot of browsing can use up a plan's monthly allowance quickly, so it suits bounded jobs better than always-on automation.
5. Claude Code (Anthropic)
Coding agent: terminal, IDE, desktop app and browser
Claude Code is an agentic coding tool that reads a codebase, edits files, runs commands and works with git. You describe a feature or paste an error, and it plans the change, edits across multiple files, runs the tests and can open a pull request. It runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web, and it connects to external tools through the Model Context Protocol. Repository-level instructions in a CLAUDE.md file, hooks and skills let teams standardise how it behaves, and it can run in CI for review and triage.
6. Codex (OpenAI)
Coding agent included with every ChatGPT plan
Codex is OpenAI's coding agent. It runs as a CLI, an IDE extension, a cloud service that works in isolated containers, and inside the ChatGPT apps, and it can be connected to GitHub to open pull requests. It takes a task, gathers context from the repository, makes the change and runs what it needs to verify it. Usage is metered per plan: OpenAI describes Plus as enough for a few focused coding sessions a week and Pro as enough for a full week across projects. The model you pick and whether a task runs locally or in the cloud both affect how far a plan stretches.
7. Microsoft Copilot Studio
Workflow builder: governed agents for Microsoft 365 and Teams
Copilot Studio is Microsoft's platform for building and publishing agents that run in Teams, SharePoint, Microsoft 365 Copilot and external channels. You build with natural language or a graphical designer, ground the agent in company data, and give it actions through connectors to Dynamics 365, Power Platform and third-party systems. Usage is metered in Copilot Credits, which replaced the older per-message unit in September 2025; the credits an action consumes depend on its complexity. Tenant controls, audit and data policies are the reason enterprises pick it, and also why setup takes longer than with consumer tools.
8. Zapier Agents
Workflow builder: agents across Zapier integrations (formerly Zapier Central)
Zapier Central has been renamed Zapier Agents. Where a Zap is a fixed sequence of steps, an agent is given instructions, tools and knowledge sources, and decides which connected apps to use for a given trigger or chat request. It can browse the web, look things up in attached documents and run actions across Zapier's large integration library. Usage is measured in activities rather than tasks, so Agents usage does not draw down a Zap plan. The tradeoff is the same as with Zaps: the agent can only do what an integration exposes.
9. Lindy
Workflow builder: Slack-first AI teammate for recurring office tasks
Lindy positions itself as an AI teammate that lives in Slack and connects to Gmail, Notion, HubSpot and other tools, plus MCP servers. You set up agents that run on a trigger (a new email, a calendar event, a Monday morning schedule) and carry out a defined job: triage the inbox, update the CRM, transcribe and summarise a meeting, or prepare a report. Its strength is the low setup cost for common office workflows. It is credit metered, and agents with many steps use credits faster than the plan names suggest.
10. n8n
Workflow builder: open source, with agent nodes and free self-hosting
n8n is a visual workflow automation tool with first-class AI agent nodes: you drop in a model, give it tools, memory and a system prompt, and wire it into integrations and webhooks. It sits between a no-code tool like Zapier and a code framework like LangGraph. The Community Edition is free to self-host, which makes it popular with technical teams who want their automations and data on their own infrastructure. Cloud plans are priced per workflow execution regardless of how many steps a workflow contains.
11. Relevance AI
Workflow builder: low-code specialist agents for sales and operations
Relevance AI is a platform for building specialist agents for sales, customer success, marketing, HR and operations, with templates, evaluation tooling and integrations to common business apps. It is low-code rather than no-code: non-developers can assemble an agent, but getting reliable results still takes care with tools and prompts. In 2026 it split billing into Actions (what agents do) and Vendor Credits (model and tool costs), and it retired its free plan, so you now need a paid plan to use it at all.
12. AutoGPT Platform
Workflow builder: the 2023 open-source agent, rebuilt as a visual platform
AutoGPT started the autonomous agent wave in 2023. The original script now lives in the classic directory of the repository; the active product is the AutoGPT Platform, a visual builder with a dashboard and a marketplace of prebuilt agents, plus a chat-driven mode for describing a workflow in plain language. The platform code is under the Polyform Shield licence, which allows personal and internal business use but not selling it as a competing hosted service; the classic code stays MIT. Self-hosting with Docker Compose has no licence fee, but you bring your own model keys and infrastructure.
13. LangGraph
Framework: durable, stateful agents with human checkpoints
LangGraph, from LangChain Inc, is a low-level orchestration framework and runtime for long-running, stateful agents. You model an agent as a graph that mixes deterministic steps with model-driven ones, and the runtime gives you durable execution that resumes after failures, human-in-the-loop checkpoints where a person can inspect and change state, and both short-term and long-term memory. It is MIT licensed and usable without the rest of LangChain. The learning curve is real, but it is the framework many teams reach for when an agent has to survive a restart or wait days for an approval.
14. CrewAI
Framework: multi-agent crews with roles and tools
CrewAI is an MIT-licensed Python framework for multi-agent workflows. Its mental model is a crew: several agents with roles, goals and tools that hand work between them, plus flows for more deterministic orchestration. It is quicker to get a first result with than LangGraph and maps naturally to business processes that already look like a team. The company sells CrewAI AMP, a managed control plane with tracing, deployment and governance, aimed at enterprises.
15. OpenAI Agents SDK
Framework: minimal, provider-agnostic agent library
The OpenAI Agents SDK is a small MIT-licensed Python framework with very few abstractions: agents are models with instructions and tools, handoffs let one agent delegate to another, guardrails validate inputs and outputs, and sessions keep working memory across a loop. Despite the name it is provider-agnostic, supporting OpenAI's Responses and Chat Completions APIs as well as many other model providers. It is a good choice when you want tool calling and multi-agent handoffs without adopting a large framework.
The Four Kinds of AI Agent
Most confusion about agents comes from comparing tools that are not trying to do the same thing. These four kinds overlap at the edges, but knowing which one you need narrows the choice quickly.
Browser and computer-use agents
These operate a browser or a whole desktop the way a person would: look at the screen, click, type, scroll. ChatGPT agent mode, Claude in Chrome, Gemini auto browse and Manus are the consumer versions; the Claude computer use API is the developer version. They need no integrations because they use the same interface you do, which also makes them slower and more fragile than an API call. Good for tasks on sites that have no API and for one-off jobs. Expect to approve sensitive steps.
Coding agents
Claude Code and Codex read a repository, make changes across files, run tests and open pull requests. Code is the domain where agents have advanced fastest, because the environment is text, the feedback (tests, compilers, linters) is immediate, and mistakes are reversible with version control. If you write software, this is the kind of agent to try first. Our AI coding tools guide covers the wider field, including editors and review tools.
Workflow builders
Copilot Studio, Zapier Agents, Lindy, n8n, Relevance AI and the AutoGPT Platform let a team define an agent with a trigger, tools and instructions, then run it on a schedule or in response to events. The scope is deliberately narrow, which is what makes them dependable: an inbox triage agent only triages the inbox. They are the right choice when you know exactly which process you want automated and you want it running unattended.
Frameworks
LangGraph, CrewAI and the OpenAI Agents SDK are libraries for engineers who are building an agent into their own product. They handle the loop, tool calling, state, handoffs between agents and human checkpoints, and leave everything else to you. Pick a framework when off-the-shelf tools cannot reach your data or when the agent is the product. Pick one of the other three kinds when you just want a task done.
How to Choose an AI Agent
Match the tool to the job, not to the demo. These are the questions that decide it.
- Do you want to use an agent or build one? If you want a task done this week, use a browser agent on a subscription you already have, or a workflow builder. Frameworks are for engineers shipping software, and they take weeks, not minutes.
- Where does your data live? Google Workspace users get the most from Gemini Spark. Microsoft 365 shops should start with Copilot Studio. If your stack is a mix of SaaS tools, Zapier Agents or Lindy reach the most apps with the least effort. If you need to keep data on your own servers, n8n or a framework is the honest answer.
- Is the task code? Then use a coding agent. Claude Code if you are on Claude, Codex if you are on ChatGPT. Both are included in plans you may already pay for, so the cost of trying is low.
- How will usage be metered? Almost every tool here charges by credits, activities, requests or executions rather than by seat alone. Before committing, run one real task and check how much of the monthly allowance it used. A plan that looks cheap can run out in a day on a browsing-heavy job.
- Can you undo what it does? Start with tasks where a wrong action is cheap: drafting rather than sending, filling a cart rather than paying, opening a pull request rather than merging. Expand the agent's permissions only after it has been reliable on the safe version.
For a general method that applies to any AI purchase, read our buyer's guide to choosing AI tools.
AI Agents vs AI Assistants: What Is the Difference?
Vendors use the two terms loosely, and some products are both, so it helps to be precise about what each word means.
An AI assistant (ChatGPT, Claude or Gemini in ordinary chat) responds to prompts. You ask, it answers; you paste text, it transforms it. It produces output that you then act on, and nothing changes in the world until you do. Our ChatGPT alternatives guide compares the main assistants.
An AI agent receives a goal and takes actions to reach it: it browses, clicks, fills forms, calls APIs, edits files and sends messages, checking the result of each step before choosing the next. The defining feature is the loop, and the defining risk is that the loop acts on your behalf with your access.
In practice the same model powers both. ChatGPT is an assistant in chat and an agent in agent mode; Claude is an assistant in the app and an agent in Claude Code or Claude in Chrome. The difference is not the intelligence of the model but the tools it has been handed and the permission it has been given to use them.
What Agents Still Get Wrong
The tools above are usable, but none of them is hands-off yet. Four failure modes come up across every category.
- Prompt injection. A web page, email or document the agent reads can contain instructions, and the model may follow them. This is why browser agents confirm before purchases and sending messages, and why you should not give an agent access to accounts it does not need for the task.
- Fragile interfaces. Sites that block automation, require CAPTCHAs or use unusual logins defeat browser agents. An API integration, where one exists, is more reliable than screen control.
- Cost surprises. Credit-metered plans reward short, well-scoped tasks and punish open-ended ones. Long research or browsing jobs can consume a month of credits in a sitting.
- Confident wrong answers. An agent reports that a task is done in the same tone whether it succeeded or not. Check the output until you have a track record for that specific task, and keep a human step in anything that sends money or messages.
Sources
Product details and prices in this guide were checked against the following vendor pages in September 2026. Prices change; treat the vendor page as authoritative.
- https://learn.chatgpt.com/docs/pricing
- https://learn.chatgpt.com/docs
- https://help.openai.com/en/articles/11752874-chatgpt-agent
- https://claude.com/pricing
- https://claude.com/chrome
- https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool
- https://code.claude.com/docs/en/overview
- https://gemini.google/subscriptions/
- https://support.google.com/gemini/answer/16596215
- https://support.google.com/chrome/answer/16821166
- https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus
- https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio
- https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing
- https://zapier.com/pricing
- https://help.zapier.com/hc/en-us/articles/26559132765325-How-is-Zapier-Agents-usage-measured
- https://www.lindy.ai/pricing
- https://n8n.io/pricing/
- https://relevanceai.com/docs/get-started/pricing
- https://agpt.co/pricing
- https://github.com/Significant-Gravitas/AutoGPT
- https://docs.langchain.com/oss/python/langgraph/overview
- https://www.langchain.com/pricing
- https://github.com/crewAIInc/crewAI
- https://www.crewai.com/pricing
- https://openai.github.io/openai-agents-python/
Frequently Asked Questions
What are AI agents and how do they work?
An AI agent is software that is given a goal and works toward it by taking actions: opening web pages, running code, calling APIs, editing files or sending messages. It runs a loop: look at the current state, decide on an action, carry it out, check the result, and repeat until the goal is met or it needs a person. A chatbot answers a question; an agent does the task. The same underlying models power both. The difference is the tools and permissions wrapped around the model.
Are AI agents safe to use?
Safe enough for low-stakes tasks if you use the guardrails. Consumer agents such as ChatGPT agent mode, Claude in Chrome and Gemini Spark ask for confirmation before purchases, sending messages and submitting forms, and let you take over. The main risks are prompt injection (a web page or email that instructs the agent to do something you did not ask for), over-broad access to accounts, and runaway usage costs. Start with tasks where a mistake is cheap to undo, give the agent only the accounts it needs, and read what it did before you rely on it.
Are there free AI agent tools?
Yes. Codex is included with the free ChatGPT plan. Manus has a free tier with daily credits, Zapier Agents has a free plan with 400 activities a month, and CrewAI offers a free Basic plan on its hosted platform. LangGraph, CrewAI, the OpenAI Agents SDK, n8n Community Edition and the AutoGPT Platform are open source and free to self-host, though you pay for the model API you connect. Lindy and Relevance AI no longer offer free plans. Browse all AI tools on AI Registry to compare.
What happened to ChatGPT Operator, Zapier Central and AutoGPT?
Operator was folded into ChatGPT as agent mode in 2025 and its standalone site closed. Zapier Central was renamed Zapier Agents. AutoGPT still exists, but the original 2023 script has been superseded by the AutoGPT Platform, a visual builder with hosted plans and a self-hosting option.
Will AI agents replace human workers?
Agents automate specific, well-defined tasks rather than whole jobs. They do well at repetitive work with clear success criteria: data entry, form filling, research, triage and routine code changes. They do less well with judgement calls, novel situations and anything where a wrong action is expensive. In practice they change what a role spends time on more than whether the role exists.
Find AI Agent Tools on AI Registry
Compare agent platforms, read reviews, and find the tool that fits the task you want automated.
