AI Agents for Business in 2026: OpenAI Agents SDK, Claude Agent SDK and n8n Compared

An AI agent is not ChatGPT in a browser tab. It is a model with access to tools and a loop around it: the agent reads the request, decides which tool to call, looks at the result and repeats until the task is done or a person has to step in. The tools are whatever you connect: your inbox, a Google Sheet, a CRM, an invoice template.

That loop is what makes agents useful and what makes them risky. A chatbot that gives a wrong answer wastes a minute. An agent with write access to your CRM can create fifty wrong cards before anyone notices. So the questions worth asking are less “which model is smartest” and more “who will set this up, what can it touch, and who approves its actions”.

# OpenAI Agents SDK: what’s inside?

OpenAI released the Agents SDK in March 2025 and describes it as a production-ready upgrade of Swarm, its earlier experiment with agents. It is a library, not a hosted service: a developer writes the agent in Python or TypeScript (the npm package is @openai/agents) and runs it wherever the rest of the product runs.

The SDK is built around a few primitives:

• Agents: a model with instructions and tools, meaning functions it may call, such as “look up the client in the CRM”.
• Handoffs: one agent passes the task to another that specializes in it, for example from triage to billing.
• Guardrails: checks on the agent’s input and output that stop a run when something is off.
• Sessions: memory that keeps the context of a conversation across turns.
• Tracing: a record of every step, which you will need the first time an agent does something odd.
• Human in the loop: built-in points where a person approves the next step.

It suits a business that already has a developer and wants an agent inside its own product or internal system, especially when the work splits naturally between several roles. It is overkill if all you need is “read the email, add a row to the sheet”.

# Claude Agent SDK and the Claude API

Anthropic offers two routes. The Claude Agent SDK (it was called the Claude Code SDK before the rename) gives you the same agent loop, tools and context management that power Claude Code, as a library for Python and TypeScript. Out of the box the agent can read and write files, run commands and search the web. You add hooks that run your code at key points, subagents for focused subtasks, external tools through MCP (Model Context Protocol), sessions, and permissions that decide which tools run automatically and which wait for approval.

The second route is the plain Claude API with tool use, where the developer writes the loop. It takes more work and gives more control.

The Agent SDK fits jobs where the agent has to handle documents and files by itself: go through a folder of contracts, extract the payment terms, put them into a table. Its permissions model matters here. Set it so that anything that changes data outside the agent’s working folder needs a person’s approval.

# n8n AI Agent node: agents without code

In n8n an agent is a block in a visual workflow. You connect a chat model (OpenAI, Claude and others are available) and at least one tool, and the agent decides which tool to call. A memory sub-node keeps the conversation history: the Simple Memory node is enough to start, though n8n advises against it for busy production workflows running in queue mode.

The tools are ordinary n8n nodes: HTTP requests, Google Sheets, Gmail, Telegram, CRMs such as Pipedrive and HubSpot. That is n8n’s real advantage. Most small-business agents are 80% integration and 20% intelligence, and n8n has the integrations ready.

Two ways to run it:

• Self-hosted (Community edition): free when only you or people in your organization create and edit the workflows. The Sustainable Use License does not let you host n8n as a service where clients build their own workflows, or resell it under your own brand. Building automations for a client is fine as long as the client cannot edit them.
• n8n Cloud: Starter at €20/month billed annually with 2,500 workflow executions, Pro at €50/month with 10,000. There is a free trial with no card required.

Where n8n struggles: long chains of agents that loop back to each other. You can build them with sub-workflows, but they get hard to read fast. If you are heading there, it is time for a developer and an SDK. We compared n8n with its closest alternatives in n8n vs Make vs Zapier.

# What does an agent cost per month?

Two costs: the platform and the model tokens. Here are the vendors’ standard API prices per 1M tokens, checked on October 22, 2026:

A worked example. An agent qualifies 100 incoming requests a day, about 3,000 a month. Each run reads around 1,500 tokens (the message, the instructions, your qualification rules) and writes around 300. That is 4.5M input and 0.9M output tokens a month:

• on gpt-6-luna or Haiku 5.5: 4.5 × $0.10 + 0.9 × $0.50 ≈ $0.90;
• on gpt-6.1-sol or Sonnet 5.5: 4.5 × $2 + 0.9 × $10 = $18.

Start on a cheap model and move up only for the steps where it makes mistakes. For most small businesses the platform, not the tokens, is the larger line in the budget.

# Three jobs agents do well in a small business

Lead qualification. A request arrives by email or Telegram. The agent extracts budget, deadline and service type, compares them with rules you keep in a Google Sheet, and either creates a CRM card with a task for a manager or sends a polite reply with materials. Write the rules as numbers (“budget under $500 and start later than three months: cold”), not adjectives. Vague prompts are the most common reason an agent misclassifies.

Invoices and documents. When a deal moves to “Won”, the agent takes the client details and amount from the CRM and fills an invoice template, for example in Google Docs. A person checks the amount and bank details and sends it. Never let the agent send invoices on its own: a wrong amount sent automatically is a reputational and legal problem. If you issue fiscal receipts through a software cash register (PRRO), keep that step manual or behind explicit approval too. More on automating paperwork for sole traders: document automation for sole traders in Ukraine.

Support triage. Messages from Telegram, email and the website go into one workflow. The agent sorts them into “delivery status”, “billing”, “complaint” and “other”. It answers routine questions from your FAQ and hands the rest to an operator with a short summary. Start with one category, such as delivery status, check a hundred answers by hand, and only then widen the scope.

# How to start in one day

1. Accounts (30 minutes). Sign up for n8n Cloud or install n8n on your own server. Create an API key with OpenAI or Anthropic and add a small balance for testing.
2. The first agent (2–3 hours). Create a workflow with a trigger (a Telegram bot or Gmail). Add the AI Agent node, choose a cheap model, and write a system prompt with examples of correct and incorrect answers. Add Simple Memory and one tool, for example a Google Sheets node with your price list. Run it on test messages.
3. Connect the business (1–2 hours). Add your CRM node or a Google Sheet as a simple CRM, and a Telegram notification to the manager when the agent hands a lead over.
4. Review before trusting it. Read the first hundred runs. Fix the prompt where it fails. Only then give the agent the next task.

# Where agents go wrong

• Too many permissions on day one. Give read access first, write access to one place second, and anything involving money only behind approval.
• No logs. If you cannot see what the agent did and why, you cannot fix it. The OpenAI SDK has tracing built in; in n8n every execution is saved in the workflow history.
• Personal data. An agent that reads customer messages processes personal data, and Ukraine’s Law “On Personal Data Protection” applies. Send the model only the fields the task needs, and check where your provider stores the data.
• Automating everything at once. One process, two or three tools, a precise prompt. Then the next one.

# See also

• n8n vs Make vs Zapier 2026: which automation platform to pick
• Custom GPT for business: a setup guide
• AI email automation: templates and auto-replies via ChatGPT
• Airtable for business in Ukraine: build a CRM in under an hour
• Best CRM platforms in Ukraine for e-commerce