Hana Field: I've had the strangest week — I kept reading press releases and feeling like I was losing my grip on language, and I think I finally figured out why.
Iris Holm: Press releases will do that. What broke it?
Hana Field: Four products launched inside one month — OpenAI Presence, Meta Business Agent Platform, NVIDIA and ServiceNow's Project Arc, Google Gemini Enterprise — and every single one of them used the words 'agent platform.' And so I started trying to work out what that actually means, and I realized, you know, it means something completely different in each press release.
Iris Holm: Right. Because they're not competing. They're not even in the same market.
Hana Field: That's the take you want to open with, isn't it.
Iris Holm: Frankly, yes. Meta is chasing distribution — WhatsApp and Messenger reach. Google Gemini Enterprise is about data gravity, keeping everything inside its own cloud. Project Arc is workflow orchestration. And OpenAI Presence — this one really got me — it launched July 22nd and it is explicitly not self-serve. It requires Forward Deployed Engineers inside each client. That is a managed services business. Someone called it a platform and nobody pushed back.
Hana Field: Hold on — explicitly not self-serve at launch, those are their words?
Iris Holm: That's the structure of it. And the question that doesn't go away — if the category leader can't self-serve, what does that say about whether 'agent platform' describes anything real at all.
Hana Field: Okay but — before we go further into whether Presence is a platform or a consulting firm with good branding — I want to back up one layer, because I think the 'agent' part is doing even more work than the 'platform' part. Like, what even is an AI agent, in plain terms.
Hana Field: Think of a regular generative AI as a vending machine. You put in a prompt, you get an output. Done. An agent is more like a contractor — you hand them a goal, they figure out the steps, they make calls, they check the results, they come back when it's finished. Autonomous, multi-step, minimal hand-holding at each stage. That's the actual definition. The MIT Sloan Management Review framing, the OECD AI Policy Observatory framing — they both land basically there: perceives, reasons, executes, with minimal human intervention per step.
Iris Holm: And a platform should be the jobsite. The thing that lets you hire and manage multiple contractors.
Hana Field: Right — and here's where it breaks down, because each of these four 'jobsites' is actually a different industry. Meta Business Agent Platform is essentially a messaging app that added automation. Google Gemini Enterprise is a cloud ecosystem keeping your data inside its own walls. NVIDIA and ServiceNow's Project Arc is workflow orchestration — closer to IT infrastructure. And OpenAI Presence is, I mean... it's a consulting firm that learned to say API.
Iris Holm: No shared integration standard across any of them. That's not a competitive market — that's four parallel experiments that happen to share a word.
Hana Field: And so when an enterprise buyer sits down to compare them — actually, no, that's the problem, you can't compare them. Distribution versus containment versus orchestration versus data gravity. Those aren't different answers to the same question. They're different questions entirely.
Iris Holm: The label is shared. The thing being built is not. That's the whole episode in one sentence.
Hana Field: And the structural fault there — it shows up in something really specific, and I want to sit with it for a second, because I think this is where you actually get your win. The Forward Deployed Engineers aren't a footnote in how OpenAI Presence works. They're the product. Without them on-site, inside the engagement, the thing doesn't run. That's not an API key. That's a staffing model.
Iris Holm: That's the structural tell. A self-serve platform scales because your engineers do the work. This scales because theirs do.
Hana Field: And so imagine — you're a procurement officer, you've just gotten budget approval to evaluate three agent platforms. You set up the calls. First vendor is Meta Business Agent Platform, which is, you know, basically WhatsApp with automation bolted on. Second is Project Arc from NVIDIA and ServiceNow — that's IT workflow orchestration, that's a whole different buying committee, that's infrastructure. And then OpenAI Presence, and you get to the contract terms and realize you're not licensing software, you're... contracting for embedded personnel. Those are three completely different vendor relationships. Same label.
Iris Holm: Three different buying committees.
Hana Field: At minimum. And no shared definition of what she was even evaluating when she started.
Iris Holm: Now here's the implication I actually find uncomfortable — OpenAI's real moat, if this model holds, is services margin and governance assurance, not scalable software infrastructure. That's a fundamentally different business. And if the market leader is running that play, what does that tell us about whether autonomous agents are actually solvable at scale without human oversight baked in at every engagement?
Hana Field: Oh — yeah. That's the uncomfortable read.
Iris Holm: And governance language is everywhere now — guardrails, audit trails, approval workflows, Google Gemini Enterprise uses it, Project Arc uses it, Presence uses it — but no independent body has audited whether any of these architectures are actually equivalent. They've converged on the vocabulary. The architecture is a black box.
Hana Field: It's four airlines all saying 'safety procedures' and assuming that settles it.
Iris Holm: And the part that makes all of this worse — which we should get to — is that buyers are locking in right now, before regulators or the OECD AI Policy Observatory have landed on any shared definition. The confusion is survivable. The lock-in might not be.
Hana Field: And lock-in is the part that's actually darker than the category confusion, because — Washington policymakers started formal scrutiny of AI agents in mid-2026, right when all four of these launched. So regulators are watching. But watching isn't defining. And enterprise buyers aren't waiting for definitions. They're signing contracts right now.
Iris Holm: What are they actually committing to, structurally?
Hana Field: That's the trap — they don't know. And the OECD AI Policy Observatory is trying to build a common definitional framework for agentic AI, but no binding standard has come out. So the buyer signs with, say, Google Gemini Enterprise for data gravity, and eighteen months later realizes she's locked every workflow into Google's cloud ecosystem, and the standard that emerges — if it emerges — doesn't map to what she bought.
Iris Holm: That's not vendor lock-in. That's definitional lock-in. You can't even audit what you committed to.
Hana Field: And then there's the 15% number — analysts projecting 15% of work decisions made autonomously by AI agents by 2028. And I actually — wait, that number bothers me more the longer I sit with it, because 'autonomous decision' is undefined in a market where 'agent' itself is contested. You can't measure 15% of something nobody has agreed to measure.
Iris Holm: The projection assumes adoption. But fragmentation may be the brake.
Hana Field: Exactly that — and there's one more thing that nobody's public governance language is touching yet. Agent memory stores. The interactions, the retrieved facts, the preferences these systems accumulate — research shows those stores are vulnerable to membership inference attacks. Real data exposure. And every platform is leading with guardrails and audit trails, and I'll give them that, that convergence on necessity is real — but none of their public positioning addresses memory vulnerability specifically.
Iris Holm: So the governance rhetoric is table stakes now — I'll grant that — but it's converging on the wrong surface.
Hana Field: The calibrated take, I think, is this: the category's fictionality isn't the emergency. The emergency is that the fiction has a deadline — buyers locking in now, before regulators or the OECD lands a definition, are making irreversible commitments to something they genuinely cannot evaluate. That's not a branding problem. That's structural.
Iris Holm: Okay, 'fiction' was probably too strong. I'll walk that back — slightly. Four real products. Fake category name. Four different novels with the same cover art.
Hana Field: And you know, I keep thinking about where we started — you couldn't read a press release without losing your grip on language. And that's still true. Except now I think the language problem isn't an accident. It's load-bearing. 'Agent platform' has to mean everything because if it meant something specific, three of these four products wouldn't qualify.
Iris Holm: The real question for any enterprise buyer in the next twelve to eighteen months isn't which one to choose. It's whether you can afford to be wrong before anyone agrees what you're choosing.
Hana Field: Sharp stop. Yeah.
Iris Holm: Good talk. Unsettling, but good.
