Manus got acquired by Meta for over $2 billion. Claude Cowork dropped and the "SaaSpocalypse" wiped $285 billion off software stocks. OpenClaw went viral and became the most-starred software project on GitHub.
That was just Q1 of 2026.
I've been using AI tools for three years now. ChatGPT Plus subscriber since day one. Over a million tasks automated on Zapier. Built voice agents, WhatsApp pipelines, a content engine plugin on Claude. Taught myself to code somewhere along the way.
And I still feel lost sometimes. If you feel the same way, that's not you being slow. It's just how fast this space moves.
Here's my observation. I think there are three types of AI tools, and most of the confusion comes from treating them as one.
Category 1: AI Chatbots
ChatGPT, Gemini, Copilot, Perplexity
You open a chat, ask a question, get an answer. The most common way people use AI today.
- Good at: Fast thinking. Brainstorming, rewriting, explaining, quick research. No setup, no learning curve.
- Bad at: Remembering you. Every session is mostly a blank slate: no persistent memory, no scheduling, weak connections to your other tools. You're the one holding all the context.
Category 2: Workflow Automation (with AI functions)
Zapier, N8N, Make, Activepieces
No chat. You build a flow: when this happens, do that, then do that. Newer versions let you drop AI steps in: "summarise this document" can sit between "receive webhook" and "update database."
- Good at: Repetitive, predictable work. Runs on schedule. Connects to hundreds of apps. Cheapest option at scale.
- Bad at: Each run is stateless. And maintenance is painful. I've built 30-step flows that worked beautifully until the business changed and the whole thing collapsed like a house of cards.
Category 3: Agentic AI / AI Agents
Claude (Cowork / Claude Code), Manus, OpenClaw
Instead of a chat or a flowchart, it's a workspace. You give the AI access to a project folder (files, instructions, context, tools) and a goal. It figures out the steps.
- Good at: Work that needs judgment AND memory. Context lives in a folder, not a disappearing chat. It can use multiple tools in sequence without you orchestrating each step. Set up Skills (reusable instruction sets) and the whole team gets consistent output.
- Bad at: Security. More expensive per task. Less predictable: sometimes it takes a weird path. The ecosystem is young.
Where this goes
The models are already smart enough. I don't think we need a breakthrough in how LLMs work for things to get dramatically better. What's lagging is everything around them: the engineering, the plumbing, the boring stuff.
I think agents are going to absorb the other two categories. Not replace them: absorb them.
You'll still need workflows. But not a human building and maintaining them. An agent generates the flow, deploys it, monitors it, fixes it when the business changes.
What I've been noticing is that the industry is solving the bloat problem with toolkits. Plugins, skills, scripts, commands. Different names, same idea. Instead of cramming everything into one massive prompt, you give the agent a shelf of tools. It pulls what it needs, when it needs it.
That's what I'm watching now. Not the next model launch. Not the next trendy demo. The plumbing: context management, data connections, memory, privacy. The stuff that decides whether I'd actually hand over my real work.
The models are smart enough. The question is who builds the best home for them.
Originally published on LinkedIn. Want to discuss this further? Get in touch.
