The Realistic MCP Situation
Super smart and well-respected business leaders are acting like and promoting MCP-style connections between AI models and third‑party software, as if “universal agents” are working and doing great work.
I don’t buy that, at least not at the scale and smoothness people are promoting and often still imagining.
Look a little closer at the powers in control.
On paper, the story being sold is simple: one model, one protocol, a bunch of tools, and your agent(s) can do anything across the internet; they can negotiate, one-click buy, they can edit video, podcasts and create a whole marketing strategy...
I am confident in agents and in agentic in some cases, but in others we have to be realistic and know the constraints in the system.
In practice, the hard part isn’t wiring tools together.
It’s incentives, control, and friction.
The merchants and platforms will defend their moats.
The idea that Amazon, airlines, banks, or big retailers will happily let headless agents make decisions, crawl interfaces, and bypass their crafted funnels is naive and many are blocking access already.
They worry about fraud, abuse, compliance, liability, margin, and who owns the customer relationship and rightly so.
That translates to rate limits, CAPTCHAs, bot detection, extra verification, and “suspicious activity” flags whenever automated behaviour shows up. The universal buyer‑agent is running into a wall long before it fills your cart.
The more agents are used then abused and agent swarms attack, the longer the battle for control is.
Software companies want to control the experience, not be reduced to a background function.
Great Product teams obsess over UX, onboarding, pricing flows, and how features are sequenced. Handing the front‑end over to a generic agent means giving up narrative, differentiation, and upsell.
So they will push key actions back into their own UI, keep the most valuable capabilities out of public APIs, and occasionally design flows that are confusing or brittle for external agents. The result is partial automation: a model can start a task, but the human still has to click, confirm, and rescue it when things go off‑script.
It’s IF! There is the reality that everything becomes conditional.
If you have the right tokens, if the scopes are correct, if your organisation’s policies allow it, if the connector is maintained, if the vendor hasn’t quietly changed the API.
The latest demos assume the integration already exists and is stable. In production, credentials expire, endpoints move, rate limits tighten, and someone has to keep the whole mesh healthy.
That operational drag is invisible in the hype cycle and painfully visible in real teams.
You Now Become The Editor
You become the editor. The promise being sold is “tell the agent your goal and relax.”
The reality is you are setting careful instructions, checking every step, catching mistakes, and re‑running flows that fail mid‑way. For low‑stakes, “good enough” tasks, that might be fine. For anything meaningful, you’re still the quality control, not the passenger.
Many people will decide it’s still faster to just use the app.
Balancing The MCP Argument
There is a positive story here. Standards like MCP make it easier for willing tools to plug into an ecosystem instead of building custom integrations for each assistant.
Some categories will lean in: developer tools, analytics, operations software that already think in terms of APIs and automation. In those domains, agents really can string multiple systems together and feel magical.
Over time, you could imagine regulation and user pressure nudging more platforms to open up their data and actions in a structured, agent‑friendly way.
Power Wins Over Plumbing
So the honest position is not “this is all hype” or “this will instantly change everything.”
It’s that protocols solve plumbing, but they don’t solve power.
The big question is not “can we connect models to tools” but “who wants to be controllable, on whose terms, and at what cost.”
Until that shifts, we’re likely to get pockets of great agentic experiences, surrounded by a lot of guarded, human‑centric software that prefers to be the main character, not just a function call.
Here are some interesting takeaways from the Meta/Muse Cheeky Pint podcast out today about Muse and their vision for agentic
The Architecture of Personal Agents - timestamp: 9mins 30
Agentic Commerce and Future Vision - timestamp 67mins 21s
