AI Agent Platform: How to Choose Between Builders

An AI agent platform is a hosted or self-hosted product where you build an agent by configuring a model, tools and triggers instead of writing the whole system in code. As of 31 October 2026 the main choices are n8n, Zapier Agents, Make, and code-first frameworks. Flowise is archived and OpenAI Agent Builder shuts down on 30 November 2026.

The platforms side by side

The code-first alternatives are the OpenAI Agents SDK, the Claude Agent SDK, LangGraph and CrewAI. They are not platforms in the sense of this page, but they are what you pick when a platform runs out of room. See AI agent frameworks.

What a platform actually gives you

Think of an agent as four parts: a model that decides, tools it can use, a trigger that wakes it up, and a record of what it did. You can build all four yourself. A platform sells you the last three, and often a shortcut on the first.

Triggers

An agent that only answers when you open a chat window is a chatbot. The value of a platform is usually in the triggers: a new form entry, an incoming email, a schedule, a webhook, a ticket in a help desk. Zapier's agents page lists templates such as a support email responder, a GitHub pull request notifier and a candidate ranking agent, all of which depend on an event starting the run. n8n's README lists webhooks, schedules and many integrations as the usual start points.

Tools and integrations

Counts differ, and each is the vendor's own claim:

• n8n: 1500+ integrations and 9,000+ templates, per its README.
• Zapier: "9,000+ apps", per its agents page.
• Make: a library of modules and scenarios, and the Make AI Agents help index says an agent's tools can be "modules, scenarios, and MCP tools".
• Flowise: its own node library for models, vector stores and tools.

A bigger number is not automatically better. Check that the three or four apps you need are there, and look at how deep the integration goes: can it search records, or only create them?

Memory and knowledge

Most platforms let you attach documents or an FAQ so the agent answers from your material. Zapier says you can add "FAQs, docs, and public links". Make's agent settings take "instructions, knowledge, and tools". Flowise grew out of retrieval chat flows, so it has the most visible knowledge tooling of the group.

Observability

When an agent does the wrong thing, you need to see what it saw. n8n lists "human approvals and observability" in its README. Zapier has an activity view and says an agent asks before continuing once it hits its per-run cap. Ask of any platform: can I see each step, the inputs and the model's reasoning? If the answer is no, you cannot debug it.

Price models explained

Platforms bill in different units. Compare the unit before the dollar.

Zapier detail, from its pricing page: Professional plans scale from 750 tasks at $19.99 a month (annual billing) up to 2,000,000 tasks at $3,389 a month. Month-to-month billing costs more, for example $29.99 for the 750-task tier. The Team plan starts at 2,000 tasks for $69 a month annual. Agent activities are a separate unit and "do not affect your task usage", with a per-run cap of 10 activities on Free and 40 on paid plans. Zapier's blog now says agents are becoming an AI step inside Zaps, billed in tasks, so check which unit your account uses; see Zapier Agents.

Make's page states one number: the Free plan gives 1,000 credits a month. It also warns that features using Make's AI Provider "may use more" credits, with no amount given.

The hidden cost: model usage

Platforms differ on who pays the model provider. Some bundle it into credits. Some let you bring your own key. Self-hosted tools like n8n and Flowise always use your key. If you run an agent every five minutes with a long prompt, the model bill can exceed the platform fee. Work out a rough monthly run count and token size before you commit. Prices for models themselves are on LLM API prices.

Two platform stories from 2026

Flowise: archived

The Flowise repository says "Flowise has been archived." The maintainer's note in the Future of Flowise discussion says the team decided to wind down operations, citing developers moving to coding agents and the limits of rigid low-code workflows. Dates in that post: code freeze on 29 July 2026, repository archived 13 August 2026, and the end of official core-team support on 31 August 2026. The code stays under its license, and the post encourages forks. GitHub reports the license as unrecognised because the LICENSE file is a mix: Apache 2.0 for most code, with separate terms for the enterprise directory. Flowise Cloud's status was not stated in the sources we read.

What this means: if you run Flowise today it keeps working, but you get no fixes and no new model support unless a fork supplies them. Treat it as a migration candidate, not a new start.

OpenAI Agent Builder: shutting down

OpenAI's deprecation notice schedules Agent Builder to shut down on 30 November 2026. ChatKit remains. If you built there, export to the Agents SDK or another runtime. Our OpenAI Agent Builder page has the migration notes.

The pattern behind both stories is worth noting. A visual builder that depends on one company can vanish with a notice. Prefer tools where you can export your workflows as files you own.

Self-hosted or hosted

n8n's licence is fair-code, not OSI open source: source available, self-hostable and extensible, with limits on commercial reselling. If a policy at your company says "open source only", ask legal before choosing it. We cover the details on the n8n page. For private deployments across the board, see self-hosted AI and cloud vs self-hosted AI.

Platform or code

A platform is the right tool when:

• the work is connecting apps you already use
• a non-developer needs to maintain it
• the logic is a handful of steps with one or two model calls
• you value the integration library over control

Write code when:

• the agent needs custom loops, branching or long-lived state
• you have strict testing or deployment rules
• you need to run it inside your own product
• you are hitting task or credit limits

Many teams use both: a platform for the glue, a framework for the one hard agent. Our guide to creating an AI agent walks through the code route, and build vs buy AI agents covers the wider decision. A platform does not replace a coding agent either; for engineering work see AI coding agents.

Questions to ask before you pick

1. What starts the agent? If the trigger you need is missing, nothing else matters.
2. What can it touch? Give agents the narrowest permissions. A support agent should not be able to delete customer records.
3. Can a human approve? For anything that sends money, email or public posts, you want an approval step.
4. What does a run cost? Estimate tokens, tasks and credits for one run, then multiply.
5. Where does the data go? Check the vendor's data terms for your plan.
6. Can I leave? Look for export, an open licence or an API.
7. Is it still maintained? Check the last release. See the Flowise story above.

Which to pick

Common mistakes when picking a platform

• Buying on the demo. Every vendor shows a clean flow. Rebuild one of your own ugly, real processes in the free tier before paying.
• Ignoring the unit. A "task", an "activity" and a "credit" are three different meters. Write your expected monthly runs next to each meter before you compare plans.
• Giving the agent every permission. Connect each app with the narrowest account you can, and require approval for tools that send, spend or delete.
• No owner. Someone on the team must read run history weekly. An unwatched agent drifts.
• Forgetting the exit. Keep prompts, tool lists and test cases in your own repository so a closure like Flowise's costs you wiring, not thinking.

Our take

This is an editorial reading of vendor pages and repositories, not a hands-on test.

The AI agent platform market looks crowded, but the useful split is simple: platforms that start from automation (Zapier, Make, n8n) and tools that started from LLM apps (Flowise, Agent Builder). The first group is surviving; the second group lost two members in a single year. That does not prove the automation-first group is safer forever, but it has an advantage: an agent is just one more step in a workflow that already has triggers, retries and logs.

Two practical rules follow. First, pick on triggers and integrations, because that is the part you cannot easily copy elsewhere. Second, keep the agent's instructions and tool descriptions in a plain text file in your own repository. If the platform changes terms or closes, you rebuild the wiring, not the thinking.

If your plan is to ship an agent inside your own product, skip the platforms and start with a framework. If your plan is to save your team ten hours a week of copy-and-paste, start with Zapier or Make for speed, or n8n if you want to host it.

Checklist

• Write down the trigger, the tools and the approval step in one paragraph
• Confirm each tool you need exists in the platform
• Estimate runs per month and cost per run
• Check the free tier limits against a real week of use
• Decide self-hosted or hosted, and who owns the upgrades
• Save prompts and tool descriptions in your own repo
• Re-check prices and status on the vendor page the day you subscribe