When your AI agent makes life harder for you.
There are some simple yet terrifying situations you could get into with AI agents. By AI agents we mean “some kind of autonomous” bit of code that goes beyond GenAI “write me a poem” and instead says “give me some instructions” and then takes action.
This article presumes you are familiar with the concept of agents, but you don’t require a deep technical appreciation.
Listen to the blog readout by author, Steve Chambers
“What the hell do you know, Steve? 🤡”
Despite appearances, I’m not an AI luddite writing a FUD article about AI agents. I love the concept of AI agents and have been thinking about, writing about, and creating them since before they were a thing. I’ve loved all the tech epochs I’ve been through in 30 years, but I’ve never been blind to their downsides. I used to run a “Stevie Chambers Therapy Session” on “The Top 10 Problems with Virtualization” for banks in the UK, Europe, and USA. That was all about seeing the problems and avoiding them before they hit you. It was a +ve not -ve experience.
I also know the dark side of AI agents, and I want you to know too!
Where to begin? Here are three trivial yet impactful examples of what can go wrong with AI agents. If I were you, I’d raise these in any skeptical AI agent conversation (e.g. with someone selling them).
There’s no doubt AI agents are already helpful in some use cases and will get better over time.
However — and I’m sorry to be Captain Boring or Mr Obvious here — AI agents still require a qualified human to hold the reigns (human in the loop), and they also require production-quality automation wrapped around them (embedded AI agents).
You need five things to be “good” at AI agents: code, automation, LLMs, be a domain expert (e.g. sales), and common sense.
Alienating your contacts
“My AI agent co-worker just cancelled a meeting with an important opportunity with a curt message. I only found out after the meeting didn’t happen and I received an unusually cold email from a previously warm contact. Who’s fault is this?”
Deleting your data
“My AI agent co-developer just deleted files on my computer, but I only found out later when I couldn’t find something. I discovered what had happened. The LLM apologised and advised me to restore from backup. But there are no backups. Who’s fault is that?”
Building your own Skynet
“My AI network eng agent has been evolving my network switch and router configurations without issue… but now I have an issue and I don’t understand the configuration (because the AI is the “owner”). I’m not sure how to prompt the LLM to find the answer… because I don’t understand the problem. How do I fix this?”
So what can you do?
These guidelines come from my own experience and from what I’ve seen other advanced practitioners do.
You won’t find these discussed by influencers and most Youtubers, because they often just do very basic examples. They also want you to go gung-ho on AI agents and buy their course. 🤑
What you can do also depends on who you are, like we said at the top of the article (five things you need). There’s a difference between a multi-talented solopreneur with full control over all of their IT, or a leading autonomous AI agent developer, and someone working in a team, in an enterprise, who is missing some skills.
- Check out Ruv who is a leader in autonomic coding agents.
- See what John Rush is doing with business agents as a multi-SaaS solopreneur.
1. Embed agents in automation
Agents require inputs, outputs, data validation, rules, logging, and much more. They are part of a system. At the simplest level, they can be thought of as “simple leaves on a tree”. At a higher level, they can be a complete branch with sub-branches and many leaves. Some advanced folks are building a whole tree of autonomous coding agents.
You are better to use a mature automation platform for this like N8N (there are about 40 of them on the market!) and add an “AI Agent Node” or similar, so you can track what it does, how it does it, and more.
2. Put a human in the loop
It’s not advisable to leap from “human do all” to “AI agent do all”, and instead you should gradually introduce AI agents.
Begin with small specific tasks. Have a human check the input and ask (e.g. the prompt), and check the output (validate) which can include “approving the action”.
This is call Human in the Loop: the human can observe and approve/deny what the agent does.
One way HitL (Human-in-the-Loop) can be done “where you work” is by using webhooks to integrate tools like Slack where humans lurk and can review and interact with “approve” or “ndenyo” — and everyone else can see that too.
3. The best AI agent is no AI agent?
This sounds counterintuitive, but rushing ahead and “agentizing” everything from a standing start is not a wise strategy.
This desire can arise when people do a TCO/ROI model, set expectations that are too high and too soon, and therefore there’s a mad rush to use agents everywhere, prematurely.
There are sometimes non-agent answers such as traditional RPA (Robotic Process Automation) and just simple process automation. In fact, one might argue that the path to agents is:
- Understand the manual steps, have a well documented successful procedure/process.
- Automate the procedures using N8N or similar.
- If this can be improved by an AI agent, then do that (or leave it alone).
Into the AI skeptic mind
What to do with this information? The purpose of this piece is not to put you off of AI agents, or to scare you, but instead it’s to give you an insight into the AI skeptic mind.
- You won’t go out of business tomorrow through missed opportunity if you don’t do AI agents today. Turn the FOMO down to 2.
- You might go out of business tomorrow through unexpected negative impacts of AI agents today. Turn the skeptic mindset up to 10.
You won’t go out of business tomorrow through missed opportunity if you don’t do AI agents today.
Turn the FOMO down to 2.
You might go out of business tomorrow through unexpected negative impacts of AI agents today.
Turn the skeptic mindset up to 10.
More to come from me, Steve, and Viewyonder on AI, LLMs, Agents, Automation, and applying it to real world business.
Get in touch for feedback, questions, engagement.
Work with Steve and Viewyonder
Get in touch and get engaged. From a free 30 min chat to a long running retainer — we can work something out, no strings and no worries. Go to Viewyonder Services ->
Want to talk about AI agents?
You can grab 30 mins with me here:
