Creating Supervised Agents with a Human-in-the-Loop
Now Everything Starts Changing
Please Note: Once again this is a particularly long post, so subscribers will find the emailed version of this issue cut off at the end. Please go to the web version for the complete version.
If You Find AgenticMSP Valuable: Please recommend AgenticMSP to your friends and colleagues. Also, please offer up your comments, observations, questions, or any other responses in the chat below. Let’s get a hearty conversation going!
A few weeks ago, most AI analysts named Claude Code as the best possible platform for creating AI agents. Yesterday that shifted to Codex with the announcement of ChatGPT 5.5. It wasn’t so much that the large language model (LLM) itself was that much ‘smarter,’ it was a combination of the model, the harness, application access, computer use, data handling, and other capabilities that were introduced along with it.
MSP Bottom Line: It’s getting harder to keep up with the complexity and variety of changes in agent-building platforms, way beyond the size of the context window and the volume of data the model is trained on. You may choose today and feel the need to change tomorrow, and again the next day, and so on. There’s nothing easy about this.
Many developers have joined Claude-Code-Creator Boris Cherny in acknowledging that they no longer write any code themselves. Their preferred platform is writing 100% of their code for them. This has huge implications for MSPs. Those of you who have systems architecture, systems analysis, data management, information architecture or other fundamental knowledge can now seriously consider moving your business forward beyond that illusory borderline between hardware and software. You can now find yourself directly helping you customers to create value from their data and applications.
No Longer Cheap and Easy
Many people who were accustomed to enjoying access to search for free were disconcerted by the idea of paying $20/month for chatbot access. Some of those then started playing with Claude Code, OpenAI Codex, Perplexity Comet browser, and other platforms that moved them past the chatbot to use of the API, for which they had to pay for tokens. Those tokens ran out fast, and they found themselves investing more and more as they learned how to work with their platform to develop really useful software for themselves.
It was reminiscent of the early days of Azure when companies suddenly received invoices for consumption overages beyond their subscription, and those invoices were sometimes in the hundreds of thousands of dollars. That’s what we call career-altering…
We’re now back there again. The days of having our utilization subsidized silently by venture capital investments is ending. Anthropic was forced to add gigawatts of compute to resolve service delays and outages due to extreme demand. OpenAI is jumping on that in their effort to catch up. The crazy billions and trillions we’ve all read about in AI investments are starting to reveal themselves for truly being crazy.
MSPs must become adept at monitoring and measuring token consumption for their clients or be swept away by the fiscal “surprises” that will befall them.
There is No AI “Ginsu” or Swiss Army Knife
Remember the “ginsu” knife; chops, minces, slices, cuts, performs minor surgery… Like the swiss army knife, the one tool you needed for, well, everything? There’s no AI version of that.
Many entrants into the AI product market promise that theirs will be the only tool you need to create effective agents, orchestrate their activity, facilitate their data access, and integrate with literally every other tool you use.
Don’t buy it.
At this time, there is no “one-size-fits-all” AI solution. The foremost experts consistently agree that the right product or platform is the right one for each given use case, each individual application, which is nothing new in IT. The problem this time is the anticipated expense. But we’ve faced that before when we went ‘multi-cloud.’
So, You Want to Build AI Agents – Be Ready to Be the Most Important Element in the Agents You Build
In the past two issues we discussed assistive AI, namely the chatbot that answers all our questions and can converse with us about issues. We then graduated to collaborative or copilot AI in which the model worked with us to create useful tools.
Now we want to experience all the benefits Satya Nadella first heralded us with in October 2024 when he welcomed us to “The Agentic Age” at the Microsoft Ignite conference. It’s unlikely even he saw what would ensue.
