One speaker (a VC) said his number for evaluating AI-native startups is ARR per engineer, and that the number ought to be going up. Almost every talk and every booth at the AI Agents Conference was selling a fix for something that broke this year when agents hit production. Observability, governance, supervisor agents, data substrates, "someone's gotta babysit the bots."
But what's actually still going to be around in a couple years? What's defensible and durable?
The old SaaS pitch was simple. We bundle the expensive engineering investments and domain expertise into a tool. You'd pay for the tool and generate outcomes, but it would be rare for the software company to have real alignment to the actual value created from those outcomes.
That's breaking from two ends at once. In the direct-from-imagination era we're moving towards, engineering labor is approaching free. One of the most telling trends is the shift from companies bragging about the size of their engineering teams, towards how much ARR they can generate per engineer.
You can vibe-code much of what those booths were selling in a few days or weeks if you have the domain knowledge. The old software model was actually based on under-utilization; the most profitable SaaS companies are frequently those whose customers underuse it (fixed price for the customer, but variable cloud costs for the vendor).
Pricing is moving to "token markup." Maybe we'll get to 2-4x revenue for the software, because outcomes are more valuable; but margin compresses because transactional intelligence (i.e., the cost of running the LLMs that power many systems) is basically arbitraging token costs against outcome value.
So everyone on that floor was implicitly betting on a new moat to replace the old one. I'm not too confident that these will hold...
The most popular bet was on encoded domain expertise (e.g., the sales engineers at Harvey, a legal AI platform, are actually lawyers). I think this works *now* because we're still in the phase of "wow, this technology works like magic." I'm less convinced this is actually durable.
Why: Prompt architecture is text. It's portable. The expertise underneath it is often abundant (e.g., there are over a million lawyers in the USA). The righteous destiny for this category ought to be open marketplaces of prompt architecture and/or crowdsourced best-practices. Not trade secrets. The companies trying to build closed prompt moats are going to lose to open ones that iterate faster (which simply parallels the fact that much software engineering is rapidly becoming commoditized to agentic engineering and the burgeoning quantity of ready-made GitHub repos).
There are many people pursuing the data substrate; in short, this mirrors the early days of the Web when everyone scrambled to open up legacy data to dynamic standards-based Web UI. Agents will have 100-1000x the data demands of these Web apps, so it makes sense that we need tools to connect them, govern them and comply with regulatory obligations.
Newer entrants extend this further, wiring up databases, pipelines, Slack threads, and tickets into context graphs agents can reason over. As I noted above, all this still seems magical. Connect a database, watch an agent crawl the schema and produce a chatbot interface and easy-to-change dashboards.
But strip the magic away and most of these are prompt architectures on top of LLMs plus a data-ingestion layer. Once data-access standards mature (MCP is already doing this) and prompt architectures go open-source (alongside much of this wisdom increasingly getting pretrained into the LLMs themselves), that magic stops being proprietary. You'll be defending yourself against the same architecture built internally by your customer's eng team, or against an open-source version that's objectively better.
The observability incumbents: these might do better but only at Stripe-like ubiquity where trust is the overriding value (who doesn't trust Stripe at this point?). The ones who survive are probably going to fuse with the audit and compliance function rather than stay pure observability.
That's why I keep coming back to one arbitrage that seems critical: trust. This will be especially important in regulated industries, but it reminds me of the old (albeit now hilariously outdated) adage about "nobody ever got fired for choosing IBM." If your competitor can be vibe-coded over a weekend and your customer is a bank, why do they pay you 50x more? It isn't the engineering, it probably isn't even the expertise. The data plumbing will get commoditized, so it can't be that either... It's that you've shifted the risk to a third party who can actually price and defend against risk: SOC2, the named CEO who testifies in court and Congress, a legal team that takes calls, an indemnity wrapper for underwriters. Maybe this means that things actually get commodified into a financialization wrapper, rather than a way to package R&D (FinTech startups back to the front?!)
The version of this future I'd actually bet on: a commodity substrate (LLMs plus open prompt architectures plus standardized data access), topped by a thin layer of regulated insurance companies that price the risk of agent failure in compliance-driven industries. The middle layer (prompt-architecture-as-product vendors) is vulnerable to an awful lot of margin-squeeze.
Most of the floor was trying to build that middle layer.
[link] [comments]
