AI agents platform guide: What builders do and how to pick one in 2026

Plans, prices and features checked on the official pages on November 4, 2026.

An AI agents platform is software for building, running and monitoring agents: programs that use a language model to pursue a goal, call tools and decide their own next step. The right one depends on who builds the agent. No-code tools such as Zapier Agents and Lindy suit individuals and small teams, workflow tools such as n8n suit technical teams, and Microsoft, Salesforce and AWS target companies already on their stacks.

The six platforms below are priced from their official pages as of November 4, 2026. One notable fact: OpenAI says its visual Agent Builder is being shut down on November 30, 2026.

What an AI agents platform is and what it adds to a chatbot

A chatbot answers a question. An agent is given a goal, then picks tools, reads results and continues until it finishes or asks for help. Agents are built on large language models (LLMs), but an agent is not the same as an LLM: the model is the reasoning part, and the agent adds instructions, tools, memory and a loop around it.

A platform or builder supplies the surrounding pieces so you do not write them yourself: connectors to apps, a place to run the agent, memory, logs, permissions and testing. That is also the answer to “where to build AI agents”: either in a no-code builder, in a workflow tool, or in a developer platform where you bring your own code.

Agents versus agentic AI, and when an agent is the wrong tool

The terms overlap. In common usage, an AI agent is one system that completes a task with tools, while agentic AI describes the wider approach of systems that plan and act with some autonomy, including several agents working together. The two are often used interchangeably in product marketing.

Use an agent when the steps are not known in advance, such as researching a lead or triaging a mixed inbox. If the steps are fixed and repeat exactly, a plain automation is cheaper and more predictable. Several platforms below mix both: a fixed workflow with an agent step where judgment is needed.

What builders let an agent do

Across the products, the common capabilities are similar:

• Act in your apps: read email, update a CRM, post in Slack. Lindy lists 1,500+ integrations plus MCP, the open protocol for connecting tools.
• Browse and search the web: Zapier counts web browsing, scraping and search as agent activities, and AWS sells a cloud browser and web search for agents.
• Run code: AWS AgentCore includes a Code Interpreter that runs agent-generated code in sandboxes.
• Remember: AgentCore offers short-term and long-term memory; Lindy says its memory is stored in plain files you can edit.
• Make phone calls: Lindy lists phone calls as a new capability, and Salesforce prices voice actions separately. Salesforce also announced it is buying Cresta for voice agents, covered in our report on the Cresta deal.
• Learn over time: in practice this means stored memory and saved skills, not the underlying model retraining itself.

Agents can also talk to each other. Standards such as MCP and agent registries, like the AWS Agent Registry that catalogs agents, skills and tools, exist for this purpose.

Six platforms compared by who they are for

Zapier Agents: No-code, billed in activities

Zapier Agents bills in activities, not tasks, so agent work does not use your Zapier task allowance. According to Zapier, the Free plan includes 400 activities a month and Pro includes 1,500. Free runs are capped at 10 activities and paid runs at 40, after which the agent asks before continuing.

Lindy: An assistant for email, meetings and calls

Lindy describes itself as an “AI employee” that works in Slack, iMessage, Gmail and a Chrome extension. You teach it a task once or describe a recurring routine in a sentence. It offers a Free plan with $50 in credits that last 7 days, per its pricing page.

n8n: Workflow builder you can self-host

n8n counts one execution per full workflow run, however many steps it has. It publishes a Community Edition on GitHub for self-hosting, which matters for teams that need data to stay on their own servers. Its paid cloud plans are hosted by n8n.

Microsoft Copilot Studio: For Microsoft 365 organizations

Copilot Studio is sold as a tenant-wide license with capacity packs of 25,000 Copilot Credits, or through Azure pay-as-you-go. It fits companies whose data and identity already sit in Microsoft 365.

Salesforce Agentforce: Agents inside the CRM

Agentforce puts agents on Salesforce data. It offers a free Salesforce Foundations tier that includes Agentforce Builder, Prompt Builder and Agent Script, plus usage-based and per-user options.

Amazon Bedrock AgentCore: Infrastructure for developers

AgentCore is a modular set of services for running agents in production. AWS says it works with frameworks including LangChain, the OpenAI Agents SDK, the Claude Agent SDK and Strands SDK, and supports any model. For a look at one feature, see how AgentCore lets you watch a coding agent’s web app live.

Plans, prices and free options

Prices checked on November 4, 2026. Zapier annual prices are for the 750-task tier. n8n prices are the annual rates shown on its page. Lindy’s page does not state the length of a paid billing period or the exact meaning of the per-user base beyond what is listed.

How to choose and get started

1. Write down one job. Pick a single repetitive task with a clear finish, such as sorting support emails or drafting follow-ups.
2. Match the platform to where your data lives. Microsoft 365 shops lean to Copilot Studio, Salesforce shops to Agentforce, AWS teams to AgentCore.
3. Decide who builds. Non-technical staff should start in Zapier or Lindy. Developers who want control can choose n8n or AgentCore.
4. Start on the free option. Run the one job with real but low-risk data and watch the logs.
5. Set limits before scaling. Cap spend, give the agent only the permissions it needs, and require approval for actions that send, pay or delete.
6. Review results weekly. Keep the agent only if it saves measurable time.

Developers can also look at how coding agents learn a new API through skills and plugins.

Costs, security and why agents fail

Pricing units differ, which makes direct comparison hard. Zapier counts activities, n8n counts workflow runs, Salesforce counts credits or conversations, Microsoft counts Copilot Credits, and AWS bills compute and calls. A busy agent can use far more units than a first estimate suggests, so check the unit before you budget.

Agents fail for ordinary reasons: unclear goals, too many tools, poor data, and no human check on risky steps. Security is a real concern because an agent holds credentials. Security vendors are building controls for this, for example CrowdStrike’s kill switch that cuts off rogue agents at the gateway. Whatever you choose, use least-privilege access and keep logs.

Vendor changes to watch before you commit

OpenAI’s documentation says Agent Builder, its visual canvas for multi-step agent workflows, is being deprecated and is scheduled to shut down on November 30, 2026. According to the same page, ChatKit remains available, and workflow code can be run through the Agents SDK. Anyone who built on Agent Builder needs to plan a move.

Google also sells an agent platform on Google Cloud, but its official pricing page could not be read for this guide, so no Google prices are given here. Check Google’s pricing page directly if you are comparing.

Common questions