We were promised autonomous AI agents. But got Workflow Automation 2.0 instead.
2025: The Year of AI Agents 😄
2025: The Year of AI Agents 😄 2026: The Year of Workflow Automation 2.0 😐
In early 2025, I was buzzing with excitement.
I built workflows in n8n that could do prospect research and send personalized emails as if they were from me. I saw the possibility of automating entire knowledge jobs with AI agents.
“2025 is the year of AI agents”, I believed it. With a 200k to 1m token context window, models can process 500+ to 2500+ pages of text. More than enough to give them any complex instruction to follow.
Fast forward to 2026. After using Claude Code with Opus 4.5 (one of the best models available) heavily for a few months, I’ve since lowered my expectations.
At 80k token context, we got poor instruction-following, hallucinations, Claude Code telling me a task is completed when it clearly isn’t.
Here’s what I’ve learned:
What DOESN’T work: Giving AI a detailed plan and a long-running task, then expecting it to finish reliably.
What ACTUALLY works:
- Break tasks into VERY SMALL sub-tasks
- Create specialized skills for how to DO each task
- Create specialized skills for how to REVIEW each task
- Create master skills with a step-by-step workflow that spawn agents to repeat do + review cycles
Each skill? Best kept within 40k tokens.
That doesn’t sound like autonomous AI agents as promised. It increasingly feels like I’m doing workflow automation 2.0.
So what’s the difference?
Workflow Automation 1.0: → Steps drawn as flowcharts → Code written to process data at each node → Structured input required → Deterministic output guaranteed
Workflow Automation 2.0: → Steps written in natural language, as skills → Data processing described in natural language, as skills → Input can be less structured → Output is non-deterministic
We’re trading determinism and accuracy for robustness and flexibility.
That’s the real state of AI right now.
Don’t get me wrong. Workflow Automation 2.0 is still very useful. It’s just that it is a far cry from “intelligence” as claimed.
I think we’ve hit the ceiling of what LLMs can do.
I am looking forward to the next release of frontier models to prove me wrong.
What do you think? Are we in the age of AI agents, or just workflow automation 2.0?
Enjoyed this? Subscribe for more.
Practical insights on AI, growth, and independent learning. No spam.
More in AI Agents
Wan Wei, Soh "begged" me to teach this course.
Today, I am launching Foundations of Claude Code, a 4-hour course on May 7, for non-techies who want to learn Claude Code properly and walk out with their ow...
80% of GEO is still SEO.
This is what I always tell people. And the screenshot is the proof.
"How do I know if an AI agent skill is safe or malicious?"
I have been hearing this question very often recently. So I thought why not write about it.
One of my biggest AI productivity unlocks this year is the extensive use of agent skills.
In this post, I share my insights after building around 75 skills over 5 months. Coding and non-coding. LinkedIn posts, cover images, carousels, presentation...
Human will be the differentiation when everyone produces the same things with AI agents with the same skills.
Since the virality of OpenClaw educated the market about agent skills, I have seen a lot of LinkedIn posts sharing 5,678 skills covering many things that pre...
From ChatGPT to Claude Code: A Non-Techie’s Introduction to the Raw Power of AI by a Techie
Not because I don't like sharing. But because the only tool I use for 99% of my AI needs is Claude Code. And while I think it is not hard to learn, I'm never...
Wan Wei, Soh "begged" me to teach this course.
Today, I am launching Foundations of Claude Code, a 4-hour course on May 7, for non-techies who want to learn Claude Code properly and walk out with their ow...
One of my biggest AI productivity unlocks this year is the extensive use of agent skills.
In this post, I share my insights after building around 75 skills over 5 months. Coding and non-coding. LinkedIn posts, cover images, carousels, presentation...
From ChatGPT to Claude Code: A Non-Techie’s Introduction to the Raw Power of AI by a Techie
Not because I don't like sharing. But because the only tool I use for 99% of my AI needs is Claude Code. And while I think it is not hard to learn, I'm never...
80% of GEO is still SEO.
This is what I always tell people. And the screenshot is the proof.
"How do I know if an AI agent skill is safe or malicious?"
I have been hearing this question very often recently. So I thought why not write about it.
Human will be the differentiation when everyone produces the same things with AI agents with the same skills.
Since the virality of OpenClaw educated the market about agent skills, I have seen a lot of LinkedIn posts sharing 5,678 skills covering many things that pre...