Solutions
Choosing the Right Solution for Your Needs
Choosing the Right Solution for Your Needs
Written by Martin Ågren - February 14, 2026
The idea is that you need to choose a solution/technique (for creating your agent) based on your actual needs.
Example 1:
- Idea: Customer support AI chatbot for my website.
- Solution: Using AgentKit (OpenAI) to create the customer support agent and placing it on my webpage. Then using API calls to OpenAI.
Example 2:
- Idea: An autonomous AI agent that searches many webpages online, reads pdf files, and writes report to me with a reference list.
- Solution: Just activate Deep research inside ChatGPT. You do not need to build this agent, since it has already been built for you.
The list of options, on this page, is roughly sorted from the simplest to the most complex one.
Maybe an AI chatbot like ChatGPT already has the capabilities you need? Is it enough to just tell it what to do?
Agentic workflow. ChatGPT is working step-by-step towards your defined goal.
Analyzing many websites and PDF files, writing a report with a reference list.
Agentic workflow. ChatGPT is working step-by-step towards your defined goal.
Example: Gathering insights around a topic, generating a PowerPoint presentation file.
Example: Gathering insights around a topic, generating an Excel file with data.
Example, in ChatGPT: Custom GPTs / Project assistants.
Adding instructions, info, PDF files to tailor the chatbot's knowledge and behavior.
Is data security a priority? Is it not acceptable to become reliant on AI services from another country? Then local inference might be a good choice for you.
(More info about local inference here).
Read about FastBots.AI here
In short: Vibe code (text to code) a web app or website.
Example: Web app with AI capabilities. API capabilities via API calls.
If you need API calls for AI capabilities: Activate Lovable Cloud (Lovable's backend) by saying to Lovable agent: "Activate Lovable Cloud for this web app".
Default frameworks/libraries: React & Tailwind CSS.
SEO-friendly solution, since a SPA (Single Page Application) is created using React:
LovableHTML.com
Step-based workflow setup.
Note: Automation without AI is possible.
Pretty similar to Make.com but step-based (instead of module-based/node-based).
An Agent builder from Microsoft.
Example: Creating a customer chatbot for your website.
Warning: Pricey option.
Test out the tool for free.
When you want to actually use/publish your agent, you need to have a Copilot Studio message pack (minimum 200 USD per month) AND a paid Azure subscription.
Simpler than the other automation platforms. Focused on LLM agent.
Sub parts of AgentKit:
- AgentBuilder: Create the node-based workflow.
- ChatKit: Create the frontend interface for the bot.
Node 1 is the input from the user (not just any event).
Module-based (or "Node-based") workflow setup.
One of the most widely used platforms for creating AI automations and AI agents.
Note: Automation without AI is possible.
More versatile than AgentKit but also more complex.
First module/node is a trigger (which could be a webhook). Examples:
- Gmail trigger = NOT a webhook.
- Custom webhook module/node that receives data from an app.
Node-based workflow setup.
Similar to Make.com but more complex.
claude.com/download - Download the Claude Cowork app for PC/Mac, click the tab "Cowork" in the app, browse to your preferred folder for the task (that you grant it access to).
Global instructions: Under Settings >> Cowork
Folder instructions: Give specific instructions for different folders.
Paid subscription to Claude is necessary.
- Team/Enterprise: Training data disabled.
- Pro/Max: Manually disable training data: Settings >> Data Privacy Controls.
Keep the desktop app open while Claude Cowork is working. If you close the app, the session ends.
It can organize and create files and folders for you.
Generated files are delivered to the folder you chose.
Connectors: Settings >> Connectors. You can connect Claude Cowork via MCP connectors to external tools like a CRM or database.
Plugins: Add plugins. A plugin is a bundle that can contain several skills, connectors, slash commands, sub-agents.
No memory between sessions.
Popular solution for coders, from Anthropic.
Claude CODE is NOT an AI model. You use Anhropic AI models via subscription or API calls.
IDE is used, for example V S Code or Cursor.
Multi-step development - agentic workflow.
Claude Code can read and modify files inside your project folder, including editing and creating files.
OpenClaw.AI - Scroll down to Quick start, Choose PC/Mac, Copy the code and run in your terminal window...
OpenClaw is NOT an AI model.
OpenClaw is an open source system that you can run locally on your computer, and then call an AI model via API (or via subscription).
Communicate via Telegram app (or other alternatives).
Memory: The agent's memory is saved locally on your computer.
Skills: You can give it different skills (instruction files) from ClawHub.
You can let colleagues or customers communicate with your AI assistant via SMS.
Video about OpenClaw (in Swedish):
youtube.com/watch?v=ESQSzFPSQlM
Same idea as Make.com but much harder since in this case, you code the logic yourself (instead of using no-code tool like Make.com).
Retrieval-Augmented Generation - meaning the generation process is being augmented by the ability to retrieve information.
Index: RAG systems typically use an index, often a vector index. Purpose: fast retrieval from large datasets/database.
Hallucination risk is reduced by access to relevant data.
Example of an AI solution that does NOT involve a LLM.
When you need to train the AI model to classify/categorize something correctly.
Example: Supervised fine-tuning (SFT).
System prompt vs fine-tuning:
Many times, adding instructions to a system prompt is enough. But sometimes you need to fine-tune the AI model instead, to get high quality output in many different scenarios. The difference is: Fine-tuning changes the weights of the AI model, rather than just guiding it with instructions.