Agents
...and related
concepts & platforms
...and related
concepts & platforms
Table of contents
Written by Martin Ågren - October 23, 2026
AI chatbot: 💬Chat
AI assistant: 💬Chat + 🛠️Tools (for tasks)
AI agent: 💬input → 🧠🛠️Autonomous → 🧠🛠️Autonomous → Goal
AI automation: ⬭Trigger → ▭Rule → ✦AI → ▭Rule → ⬭End
Automation (without AI): ⬭Trigger → ▭Rule → ▭Rule → ⬭End
AI assistant + Agentic workflow capability = AI agent
AI + Automation = AI automation
Explanations:
"Autonomous": Autonomous decision
“Rule”: Rule-based action
“AI”: ML-based action(s)
Symbols, inspired from flowcharts: ⬭ ▭
Agentic workflow: Working autonomously toward a defined goal.
Figure: Aspects to consider when building an AI system (assistant or agent).
Infographic created by Martin Ågren (using ChatGPT).
Confusion? = Do organizations & people sometimes mean different things when they say "AI agents"? - Let's take a look at it...
According to the "AI bible" = Roughly: An agent is perceiving its environment through sensors and acting upon that environment
"AI bible" = Classic book "Artificial intelligence: a modern approach" by Stuart Russell & Peter Norvig, 4th edition (2020)
Sensors = Receive input from the environment
Environment = Physical or digital - A physical environment needs to be digitally represented for an AI agent to process it
According to IBM = Roughly: System performs tasks autonomously on behalf of user/system by designing workflow & using tools
According to me = AI system that autonomously decides & takes actions to achieve a defined goal, often using multi-step workflow & tools
- This is the type of AI agent that the AI industry nowadays typically is referring to when they use the term "agent".
"Autonomous agent" = Basically the same as "AI agent" but emphasizing the autonomy aspect.
Agentic workflow = An AI agent can operate using an "agentic workflow":
- IBM, roughly: Processes where autonomous AI agent decides, take actions, coordinates tasks with high level of autonomy.
Tools = Some AI systems can use tools - Example 1: Web search capability - Example 2: Access to Box (cloud)
Skill = Definition: Reusable capability/instructions for an AI system (for AI assistant or AI agent)
Virtual computer = Some agents runs on their own virtual computer. Such virtual computer could be a VM (Virtual Machine).
VM - Virtual Machine = The VM runs on a physical server (or physical computer). A server can host several VMs.
IBM's definition = Roughly: AI chatbot that can complete tasks (not only make conversation, which would be a chatbot)
Examples of tasks that some AI assistants can handle: Create file / Schedule appointment
Custom AI assistant = Customized version of an AI assistant - Add instructions & PDFs to tailor behavior & knowledge.
Example: Project folder assistant in ChatGPT
Demo 1, Skill = How I use a skill to generate folder names (for transactions) for my bookkeeping for my own business:
@skill-ver-folder → Upload Bank statement (kontoutdrag) → Answer questions from AI → Create folders
Demo 2, Work mode = Project assistant: System prompt + File upload + Work mode (agentic workflow)
- Prompts for the demo:
Choose chat mode: Prompt 1: --ai Prompt 2: Plugin for creating slides in ChatGPT? --www
Choose work mode: @present" & click tab ...to choose the Presentations plugin. Add this prompt:
Goal: Search briefly online about AI agents in ChatGPT, then generate only one slide about the topic.
Demo 3, Workspace agent = ⋯ → Agents → Upper right: Create: Browse templates → Share agent in the workspace team
MAS = Multi-agent system
Microsoft's definition = Roughly: Multiple autonomous agents interact to achieve individual or collective goals
Example = Some agentic coding systems can use a main agent that delegates tasks to sub-agents
Swarm - Amazon, roughly: Agents in a MAS work in parallel on a problem, decentralized (without centralized supervision)
Synonym till "AI automation" = AI Workflow Automation
AI automation = Workflow, to some degree rule-based, where one or more steps/nodes/modules use AI (e.g. an AI agent)
- Simple example: Trigger: WP post → LLM shortens text → LinkedIn post
AI component = AI automation can contain AI component(s) e.g. LLM component(s) e.g. AI agent(s) or AI assistant(s)
Automation WITHOUT AI = An automation workflow does not necessarily need to incorporate an AI component
Automation GENERATED by AI = You could create a non-AI-automation using GenAI .e.g. generating a Python script file.
Make.com uses the term "AI Automation" to describe the kind of
automated workflow that can include AI module(s),
which you can orchestrate via the Make platform.
You can see them illustrate a simple AI automation by three modules/nodes to the right:
WP post (trigger) → LLM shortens text → LinkedIn post
Upsides = Possible upsides with agents: Saves time & mental energy - Potentially increased productivity & efficiency
Downsides = Possible downsides with agents:
- Environmental footprint = Agentic workflow tend to be rather demanding from an energy consumption perspective.
- API Costs / "Agentic usage limit" = You might reduce API costs (or costly subscriptions) by decomposing tasks for AI assistant.
- Focus switching = You might need to wait some time for goal to be reached, so you might need to switch focus between tasks.
- Cognitive dept = Over-reliance on agentic use could possibly lead to cognitive dept
Alternatives to agents: Chat with custom AI assistant / Globally defined hotkeys: that steer workflow (or predefined commands)
MCP = Model Context Protocol - A simple way for LLM based AI system to connect to services/tools - MCP client ↔ MCP server
Harness / Agent harness / Orchestration layer = Part/component of an AI agent - The part that is NOT the LLM(s).
- Sub-components of a agent harness = Data + Logic + Interface (UI + Tool calling)
Agentic loop = Repeating cycle for AI agent: Decide next action → Execute action → Observe result → Repeat/Proceed/Finish
- Example: Claude Code: Harness controls the looping behavior, while the LLM decides which actions to take.
Human-in-the-Loop / HITL = A human guides, reviews, corrects, approves an AI system's work during a workflow.
FastBots = FastBots.AI - Chatbot for website - These AI chatbots CAN be set up like agents
Workspace Agents in ChatGPT = chatgpt.com/agents - Only for Business plan / Enterprise plan / Edu plan
Copilot Studio (Microsoft) = copilotstudio.microsoft.com - An agent builder - Example: Create a customer chatbot for website
AgentKit (OpenAI) - platform.openai.com/agent-builder/edit - Develop LLM agent
- Framework: Agents SDK = developers.openai.com/api/docs/guides/agents-sdk - Set up agent functionality: API, tools, etc.
- Library: ChatKit = developers.openai.com/api/docs/guides/chatkit - Create the frontend interface for the bot.
Scheduled tasks in ChatGPT - Example: Get up to date news about GenAI every morning at 8 AM.
Make = Make.com = Well-known automation platform for constructing AI automations - Module-based workflow setup.
Zapier = zapier.com - Automation platform - Step-based workflow setup - Similar to Make.com
N8N = n8n.io - Automation platform - Node-based workflow setup - Similar to Make.com but more complex
__________________________________________
Dots (OpenAI) - Subscription needed: Business Premium (not just ordinary Business plan) - chatgpt.com/dots
Claude Duet (Anthropic) - Pro plan, but waitlist-based - I signed up for the waitlist: claude.com/product/claude-duet
Muse (Meta) - Free plan with usage limits, but waitlist-based, but not possible yet to join waitlist from Sweden: muse.ai/access
Grok Bot (xAI) - Pro plan needed (20 USD per month) - Download installer from: https://x.ai/bot
Perplexity computer - Pro needed (20 USD/month) - Do not confuse with Perplexity search engine - www.perplexity.ai/computer
Manus - manus.im - Gained attention in March 2025 - Free plan includes 300 daily credits - Create account at: manus.im
Work mode / CoWork mode - Agentic workflow - Gather insights around a topic, generate PowerPoint file & Excel file with data.
- In ChatGPT: At the top of the chat interface: Click the Work tab
Deep research/Research/Researcher - Research agent (feature in chatbot) - Analyzes sites & PDFs & writes report with ref list.
- In ChaGPT: Near the prompt field, click the plus icon, choose Deep research.
Sites feature in ChatGPT - For web development
Lovable = Lovable.dev - Generate web app or website - Backend: Lovable Cloud - React, Tailwind CSS, TanStack (SEO).
Replit = Replit.com
Cursor = Primarily a desktop app - But also online: cursor.com/agents
VS Code with built-in GitHub Copilot - code.visualstudio.com/download
Codex (OpenAI) - Exists in different types: New ChatGPT desktop app / Extension in V S Code / CLI / chatgpt.com/codex/cloud
Claude Code (Anthropic) - Different types: Claude desktop app / Extension in V S Code / CLI / claude.ai/code
- Released 2025 - A major force behind the transformation of software development 2025-2026.
When these harnesses use LLM(s), they become general-purpose AI agents:
LM Studio BIONIC = LMstudio.ai/download
OpenClaw / MoltBot / ClawdBot - OpenClaw.AI - Open-source - Run locally on computer - Memory file is stored locally
__________________________________________
Skills in ChatGPT - In prompt field, type: @ and then the start of the skill name, then click it, to make ChatGPT use it.
- Tip: Sometimes best to create skills manually (rather than prompting ChatGPT to create skill for you): Can be less problematic
Example: Project folder assistant in ChatGPT
System prompt in LM Studio CLASSIC with an LLM loaded into memory - LMstudio.ai/download