AI agent tools compared: frameworks, automation platforms and ready-made agents
"Agent" became the most stretched word in AI during 2025 and 2026. Vendors use it for chatbots with one tool, for automations with a model step, and for systems that work for an hour without supervision. This guide sorts the tools in this category by what they do, gives October 2026 prices, and flags the projects that are no longer safe to build on.
A working definition helps. An agent uses a language model to choose its next action, calls a tool, reads the result and decides again. The more steps it chooses for itself, the more useful it can be and the more ways it can go wrong.
Three kinds of agent tool
| Group | Who it suits | Tools in this directory | How it is priced |
|---|---|---|---|
| Developer frameworks | Engineers building agents into a product | LangChain and LangGraph, CrewAI, OpenAI Agents SDK | Free libraries, paid tracing and hosting, plus model usage |
| Workflow platforms with agent steps | Operations and IT teams automating processes | n8n, Zapier | Per workflow run or per agent activity |
| Ready-made agents | People who want a task done without building anything | Manus, AutoGPT, 11x | Subscription with credits, or annual contract |
The line between the first two groups is blurring. n8n and Zapier both now let an agent step call any of their app connectors as tools, and LangChain sells hosted deployment. The useful question is how much of the process you want written down as fixed steps and how much you want the model to decide.
Developer frameworks
LangChain and LangGraph come from the same company. LangChain 1.0 and LangGraph 1.0 shipped on October 22, 2025, and LangChain promised no breaking changes until 2.0. LangChain is the quick path: its create_agent function gives a standard tool-calling loop that works with most model providers. LangGraph is the lower level, where you draw the agent as a graph of steps. Its main selling points, in LangChain's words, are durable state that survives a server restart, persistence for workflows that run over days, and first-class pauses for human review. Both libraries are free and open source. LangSmith, the paid tracing, evaluation and deployment service, has a free Developer plan with 5,000 traces a month for one seat and a Plus plan at $39 per seat a month with 10,000 traces. From October 2026 LangSmith prices usage in LangChain Standard Units at $1 each.
CrewAI organizes work as a crew of agents with roles, such as a researcher, an analyst and a writer, each with its own tools and a manager that hands out tasks. The Python framework is open source. CrewAI's hosted platform has a free Basic plan with a visual editor and 50 workflow runs a month, and an Enterprise plan with SSO, deployment in your own cloud and a 45-day onboarding program. CrewAI says 65% of the Fortune 500 use it, a figure it does not break down.
Model vendor kits are the third option. OpenAI's Agents SDK handles handoffs between agents, guardrails and tracing, and AgentKit, introduced in late 2025, adds a visual Agent Builder and ChatKit, a chat widget you can embed in your app. There is no separate subscription; you pay for API usage. Anthropic's Claude Agent SDK exposes the same agent loop that runs Claude Code, and Anthropic said on October 7, 2026 that Max and Team plans now include monthly API credits that cover it. Vendor kits are the fastest route if you have already chosen a model provider and the hardest to move away from later.
What happened to the OpenAI Assistants API
OpenAI released the Assistants API in late 2023 as a hosted way to build agents with threads, file search and a code interpreter. When it launched the Responses API in March 2025, it said it would move the Assistants features there and retire the old API. It announced the shutdown on August 26, 2025 and switched the Assistants API off on August 26, 2026.
The migration maps old objects to new ones. Assistants became prompts, which you create and version in the dashboard. Threads became conversations, runs became responses, and run steps became items. Thread history could not be pulled from the API after the shutdown, so teams that had not exported it had to rely on their own stored copies. OpenAI has since said reusable prompt objects are also being deprecated, which is a reason to keep your agent's instructions in your own code.
Workflow platforms with agent steps
n8n is a workflow builder that you can host yourself or use as a cloud service. Its AI Agent node can call other nodes as tools, use MCP servers and pause for human approval before a tool call. n8n bills by execution, meaning one complete run of a workflow however many steps it contains, which makes long agent workflows cheaper than on per-step platforms. Cloud Starter costs 20 euros a month billed yearly, or 24 euros monthly, for 2,500 executions. Pro costs 50 euros billed yearly, or 60 euros monthly, for 10,000. The Community Edition is free to self-host under n8n's fair-code license. A polling trigger that checks for new data every five minutes uses about 8,640 executions a month by itself, so webhook triggers save money.
Zapier connects more than 9,000 apps. Its Agents product, which ran at agents.zapier.com, is being folded into AI by Zapier, an agent step inside the regular Zap editor. Zapier has not set a date to turn off the standalone product. On the old plans, Agents Free includes 400 activities a month and Agents Pro costs $400 a year for 1,500, where each search, action or web lookup counts as one activity. Inside Zaps, agent steps use the same task-based billing as the rest of Zapier, and a single run pauses for review after 75 tasks by default. The listing in this directory is still called Zapier AI Actions, which was Zapier's earlier name for letting outside AI tools trigger Zaps.
Ready-made agents
Manus, from the Singapore company Butterfly Effect, is the best-known general-purpose agent sold to consumers. It works in a cloud computer, browsing, writing code, building slides and returning files. Meta agreed to buy it for about $2 billion in December 2025, but China's National Development and Reform Commission blocked the deal on April 27, 2026, and Manus now operates independently. It has a free plan with daily credits, Pro from $20 a month with about 4,000 credits, and higher tiers at $40 and $200. Long tasks can use credits quickly.
AutoGPT went viral in 2023 as an open-source agent that looped toward any goal you set. It is now a hosted platform with a visual block builder and ready-made AI Experts for content, sales and research, coordinated by an assistant called Otto. Pro costs $50 a month, or $42.50 billed yearly, plus a credit wallet for workflow runs. You can still self-host it for free; the platform code is under the PolyForm Shield license and the original agent under MIT.
11x sells AI sales development reps, Alice for outbound prospecting and a second agent for inbound leads. In March 2025 TechCrunch reported that 11x had shown logos of companies that were not customers and that former staff described high churn, which 11x disputed. Founder Hasan Sukkar stepped aside in May 2025. 11x does not publish prices and sells annual contracts. See our sales tools guide for other options.
Projects that are archived, pivoted or uncertain
| Tool | Status in October 2026 | What to use instead |
|---|---|---|
| OpenAI Assistants API | Shut down August 26, 2026 | Responses API, Agents SDK |
| AgentGPT | Repository archived January 28, 2026; demo still runs | AutoGPT, Manus, CrewAI |
| SuperAGI | No framework release since January 2024; company sells sales software | LangGraph, CrewAI |
| /dev/agents (Dreamer) | Launched as Dreamer in February 2026; founders joined Meta in March 2026 | n8n, Zapier |
/dev/agents raised $56 million at a $500 million valuation in November 2024 to build what its founders, former Android and Stripe leaders, called an operating system for AI agents. It launched as Dreamer in February 2026 with a personal agent called Sidekick that builds other agents. In March 2026 the founding team joined Meta Superintelligence Labs, and reports said the deal did not include the platform. Treat it as experimental.
Keeping agents under control
- Limit the tools. Give an agent only the tools and permissions it needs. Read access is safer than write access.
- Require approval for actions that matter. Sending email, spending money or deleting records should pause for a person. LangGraph, n8n and Zapier all support this.
- Cap steps and spending. Set a maximum number of steps per run and a monthly spend limit with your model provider.
- Watch for prompt injection. Web pages, emails and documents can contain instructions aimed at your agent. Do not let text from outside sources trigger sensitive actions on its own.
- Log every run. Tracing tools such as LangSmith, or run history in n8n and Zapier, show what the agent did and why.
How to choose
- Write the process down first. If you can list every step, a normal automation with one or two model steps is cheaper and more reliable than an agent.
- Pick the group. Engineers building a product feature should start with a framework. Operations teams should start with n8n or Zapier. Individuals who want a task done should try Manus or AutoGPT.
- Check model portability. LangGraph, CrewAI and n8n work with many model providers. Vendor kits tie you to one.
- Run a small pilot with a step limit, a spend cap and approval on every write action, then loosen the limits as you trust the results.
Related categories: coding tools for coding agents such as Claude Code and Codex, productivity tools for assistants inside office apps, sales tools for prospecting agents, and data tools for analysis.
Frequently asked questions
What is an AI agent?
An AI agent is a program that uses a language model to decide its own next step toward a goal, calls tools such as search, a database or an app's API, looks at the result and repeats until the task is done or it needs a person. A chatbot answers one message at a time. An automation follows fixed steps. An agent chooses the steps. In practice most useful agents sit between the two, with fixed steps where the process is known and model decisions only where judgment is needed.
What is the best framework for building AI agents?
For Python and JavaScript teams that want control, LangGraph is the most widely used option, with durable state and human approval steps built in since its 1.0 release in October 2025. CrewAI suits teams that think in roles, such as a researcher agent handing work to a writer agent. Teams committed to one model vendor often use that vendor's kit, such as OpenAI's Agents SDK or Anthropic's Claude Agent SDK. For people who prefer a visual builder, n8n gives most of the same building blocks without code.
How much do AI agents cost to run?
The largest cost is usually model usage, which depends on how many steps the agent takes and how much text it reads at each step. A short task might use a few cents of model calls, while a long research run can use several dollars. Platform fees come on top: LangSmith Plus is $39 per seat a month, n8n Cloud starts at 20 euros a month billed yearly, Zapier Agents Pro is $400 a year, and Manus Pro starts at $20 a month. Set spending limits and step limits before you let an agent run unattended.
What replaced the OpenAI Assistants API?
OpenAI replaced it with the Responses API for model calls and the Conversations API for stored chat history, and shut the Assistants API down on August 26, 2026. OpenAI's Agents SDK handles multi-agent logic in code, and AgentKit adds a visual Agent Builder and an embeddable ChatKit widget. OpenAI has also said its reusable prompt objects are being deprecated, so check the current deprecation page before building on them.
Are autonomous agents like AutoGPT and AgentGPT still used?
The 2023 idea of handing an agent a broad goal and letting it loop on its own mostly gave way to narrower agents with clear tools and limits. AgentGPT's repository was archived read-only on January 28, 2026, although its hosted demo still runs. AutoGPT continued as a hosted platform with a visual builder, ready-made AI Experts and a $50 a month Pro plan, and it can still be self-hosted. Manus is the best-known general-purpose agent sold to consumers in 2026.