I have spent a lot of time lately setting up AI agents.
Not just chatting with them. Giving them jobs.
One agent researches. One writes. One packages content. One helps build things. They have names, roles, channels, tools, memory, and rules about what they are allowed to do.
That sounds cool and cutting edge until you have to spend hours figuring out why a scheduled job stopped running because the model it used no longer exists.
That is when the future starts to feel less like science fiction and more like helping your parents troubleshoot a printer over the phone.
This is the part of AI agents nobody wants to talk about.
The boring part.
The dashboards. The permissions. The OAuth tokens. The cron jobs. The logs. The routing. The approvals. The “who is allowed to send what where” decisions. The quiet little settings that determine whether your brilliant AI assistant actually does the job or silently falls over in the corner. Gross.
The new company structure (Chat GPT)
I used to think the hard part of agents would be making them smart enough, and competent enough.
That still matters. Though they’ve certainly surpassed me in that area.
Memory was another issue… but there are solutions to that. (Read my last newsletter. It needs more love.)
Lately I think the bigger problem is making them operational enough.
A chatbot is easy. But useful “automated” agents are tiny organizations.
That is what setting up Hermes Agent has taught me.
At the high level, it sounds clean. I wanted a system where Scout handles research, Scribe handles writing, Reach handles promotion, and Dev handles technical execution. A little AI team that can help me keep up with projects, content, and ideas without me manually dragging every task across the finish line.
That sentence sounds great. And they do actually do all of those things very well…
But, the actual work is much less glamorous.
Which tools can each agent use?
Where should they post?
What should they remember?
What should they forget?
When should they ask for approval?
What happens if they use the wrong model?
What happens if Gmail OAuth expires?
What happens if a job runs perfectly for three weeks and then fails because an API changed, a token expired, or a model name disappeared from the provider list?
None of that feels like “the intelligence explosion.”
It feels like IT work.
And I mean that with respect. Mostly.
Because this is where AI starts becoming real software.
Real software has permissions. Budgets. Logs. Error messages. Version changes. Access problems. Annoying edge cases. Things that worked yesterday and somehow do not work today, even though you personally did nothing wrong.
AI agents are getting pulled into that world now.
The demo version is magical. Tell the agent to research a topic, draft a newsletter, summarize your inbox, write code, build a landing page, or analyze a spreadsheet.
The real version asks follow-up questions nobody puts in the YouTube thumbnail. (I didn’t show you the first drafts of some of these news articles, even though I prompted very specifically. AI writing is getting better, but we are definitely not there yet.)
Many decisions must be made to employ a successful agent…
Can it read your email?
Can it send email?
Should it be allowed to send email?
Where does the draft go first?
Who approves it?
How do you know what it changed?
What happens if it gets halfway through and stops?
What happens if it confidently completes the wrong task… over and over and over again?
That last one is not theoretical. It’s personal.
Monitor, Monitor, Monitor (Chat GPT)
The danger with agents is not always that they fail loudly. Sometimes they fail politely. They write something reasonable, post it somewhere plausible, and move on like a tiny consultant with no consequences.
That is why the boring layer matters.
The boring layer is what turns “AI did a thing” into “AI did the right thing, in the right place, with the right permissions, better than me, and I can prove it.”
Nobody wants to read a 40-page document about workflow rules.
Unfortunately, that document might be the product. Boringgg.
Because once an agent can take action, your messy internal process becomes infrastructure.
If you want an AI agent to help with writing, sales, support, research, analytics, or operations, you have to explain the work clearly enough for software to participate.
Who starts the task?
What counts as finished?
What information can it use?
What requires approval?
Where does the output go?
What should happen when something breaks?
This used to be corporate busywork. Now it is the difference between a useful agent and a very confident intern with admin access.
And this extremely boring work grows exponentially with company size and process complexity.
AI is not killing workflows.
It is actually moving all of the human work to workflow design.
The next wave of AI will have even smarter models. Of course it will. The models will get faster, cheaper, weirder, and better at pretending they understood your vague instructions. They will get better at fooling you.
But I think a lot of the human value is going to show up around the areas surrounding models…
Task queues.
Approval flows.
Audit logs.
Spend limits.
Memory systems.
Tool permissions.
Fallback models.
Notifications when your Gmail connection dies instead of quiet failure followed by confusion.
Sound exciting and cutting edge?
Nah.
More boring IT work.
I guess Agentic AI will be exciting stuff if you enjoy reading settings pages. I am apparently becoming that person against my will.
But this is probably what adoption looks like.
The first wave of AI felt like magic because you could type into a box and get something cool back.
The next wave is less magical and more practical.
It is the part where companies, teams, and weird solo newsletter people like me have to decide what work they actually want to hand off and what risks they can tolerate. We will have to determine what process should have been written down years ago… Before agents even existed.
AI is going to force us to document all of the things that should have already been documented, but they weren’t…. Because documenting is boring.
Definitely not as flashy as a new model launch.
But this work will matter more.
Because the future probably is not one genius chatbot that does everything.
It is a bunch of agents with boring names, limited permissions, clear jobs, and a dashboard that tells you when something broke.
Less robot butler.
More help desk ticket.
Honestly, that tracks.
Welcome to IT.
Less Robot Butler, More Help Desk Ticket (Chat GPT)
🙏 Thanks for reading
🗞️Newsletter: dustincoledata.beehive.com
🌎 Site: dustincoledata.com
- Dustin