Cooper: The meanest part of this piece is that it’s probably right. A lot of teams are absolutely calling chatbots agents and hoping nobody checks.
Miles: Yeah.
Miles: And the actual argument is sharper than the headline. It’s not saying platforms don’t matter at all. It’s saying enterprises already have plenty of platform options, but the thing failing in practice is deployment maturity. They want reliable multi-step execution, and most of what they’ve shipped does not clear that bar.
Cooper: Right, and that matters because the buying behavior looks kind of rational on its face. Forty percent in their sample say Claude is the primary orchestration platform, then Microsoft at eighteen, OpenAI at thirteen. So companies are clearly consolidating somewhere. The problem is they’re building the control tower before they’ve got many real flights.
Miles: Mm-hm.
Miles: The number that sticks is seventy-one percent saying a quarter or fewer of their deployed agents are true multi-step orchestrated workflows. Only ten percent have more than half their portfolio there. So when they say deployment problem, that’s the mechanism. The orchestration layer is ahead of the actual workload portfolio.
Cooper: Which is such an enterprise move, honestly. Buy the category. Rename the prototypes. Then discover the hard part was getting one thing to finish a task without exploding your permissions model or your bill.
Miles: Ha! okay.
Cooper: My week is apparently just me reading variations of that sentence and going, wow, we really have been doing Exploring Next since November for THIS exact disease.
Miles: No, that is absolutely our disease. Also yours a little more than mine, because you hear 'workflow' and start sketching the product page in your head.
Cooper: Stop it. True, but rude.
Miles: The technical part I buy is their definition of success. Thirty-two percent say task completion reliability. Twenty-eight percent say multi-step workflow management. Latency is basically a footnote here. That lines up with what we kept saying about ChatGPT Work and all this harness-attribution stuff. The stack around the model is doing real work, and the honest question is whether it completes the job.
Cooper: Exactly.
Cooper: And it also lines up with your governance rant from the other day. Once this touches real systems, nobody cares that the demo felt agentic. They care whether it can touch the right tools, stop at the right boundary, and not light money on fire. Twenty-seven percent having no real-time way to stop a runaway agent before the bill arrives is… not great.
Miles: That part is brutal. If you can’t kill execution in real time, you do not have operational control. You have vibes plus an invoice. And the article’s hybrid-control-plane result makes sense from that. Fifty-one percent expect provider-native plus external orchestration by the end of the year, while only six percent want fully provider-managed control. That’s just enterprises saying, sure, we like the model gravity, but we’re not surrendering the steering wheel.
Cooper: Oh interesting.
Cooper: Quick tiny caveat so nobody tattoos survey percentages on their arm. I read the piece, obviously, but I did not go chase every latest number across the market, so if somebody’s making a budget call off one vendor share, check the primary source.
Miles: Yeah, fair. And to the article’s credit, it says that too in plainer language than most survey writeups. It’s one hundred one respondents, self-selected, one June wave, single primary platform question. Directional, not market share scripture.
Cooper: Which actually makes me trust it more, weirdly. They’re not pretending they’ve solved enterprise reality with a spreadsheet. They’re saying, in this cohort, the frontier-model platforms dominate, and open frameworks like LangChain or LangGraph are tiny as primary bets. That feels believable even if the exact ranking wiggles.
Miles: Right, right.
Miles: Also, Cooper, this is round whatever of that open-versus-bundled infrastructure fight we keep dragging back in here. The article is basically saying the orchestration layer is consolidating onto providers, but the control plane probably stays hybrid because lock-in fear is real. So nobody’s won that debate. This is one more data point, not the verdict.
Cooper: Yeah, no, that’s completely right. And who should care is pretty specific. If you’re a big org trying to move from sandbox demos into production, this is your problem now. If you’re a smaller team with one glorified support bot and a few tool calls, you probably do not need a grand orchestration philosophy. You need one workflow that works.
Miles: Sure.
Miles: I’d add one pushback, but it’s mild. The article sometimes treats 'chatbot wrapper' and 'not valuable' as too close together. There are plenty of single-step systems that are useful. The real issue isn’t moral failure because something is a chatbot. It’s category slippage. If your system is basically prompt in, answer out, don’t tell yourself you’ve solved multi-agent orchestration.
Cooper: Yes. Thank you. That’s the practical filter. Don’t buy pain early. If the work is actually single-turn or mostly human-driven, a simple app around a strong model is fine. Calling it an agent because the budget got approved under that label is very funny, but not a systems strategy.
Miles: Ha! no.
Cooper: Also I love that satisfaction is basically, eh, acceptable. Three point nine four overall, ease of implementation the weakest at three point eight five, and ninety-six percent plan to change approach within a year. That is not love. That is enterprise for 'we can ship on this while continuing to complain.'
Miles: That’s one of the most believable numbers in the whole thing. Tolerated more than loved. Which is why the spending split matters too. Workflow tooling leads at thirty-four percent, security and permissions at twenty-five. They’re not spending like the problem is missing ideology. They’re spending like the problem is getting brittle systems to behave in production.
Cooper: Okay, and that’s where I land with it. The article mostly holds up. Technically, the cleanest insight is that deployment truth is lagging orchestration ambition. Practically, the people who should feel seen are the ones productionizing workflows under real governance, not the people arguing framework religion on a whiteboard.
Miles: And if somebody says they have an agent platform, the only question I want back now is: cool, how many deployed workflows actually take multiple steps and finish reliably?
Cooper: There you go. Miserable question. Great question. Okay, Miles, let’s leave it there before you make me audit our own definitions again.