THE INDEPENDENT RECORD · AGENTIC AI AS A SERVICE AboutStandardsContact
GAASAGENTIC AI · AS A SERVICE
INDEPENDENT · SINCE 2026
UPDATED DAILY
NO HYPE · NO PAY-TO-PLAY
PER-TASK PRICING NOW STANDARD ● NEW BENCHMARK: 71% TASK COMPLETION ● ENTERPRISE PILOTS UP 4X ● RUNTIME FUNDING ACCELERATES ● "AGENTS ARE THE NEW SEATS" ● MARGINS UNDER PRESSURE ● THE INDEPENDENT RECORD ON GAAS
Market

Why Some VCs Are Quietly Sitting Out the Agent Hype

Not every venture investor is racing to write checks into agentic AI-as-a-service. A meaningful minority, including some of the sharpest names in the business, are holding back, and not because they're slow or skeptical of AI broadly. They're worried about thin moats, margins that get eaten by model costs, revenue that may not stick, and valuations that already price in a flawless decade. This piece breaks down the actual reasons partners give in private, what their abstention signals about the market, and what it would take to bring them off the sidelines.

By J. Okafor · May 8, 2026 · 12 min read

Table of Contents

The Quiet Abstention No One Puts in a Tweet

If you only read the headlines, you'd think every venture firm on Sand Hill Road has gone all-in on agents. The funding tracker tells a louder story than reality: nine-figure rounds for agent-orchestration startups, seed valuations that would have been Series B numbers two years ago, founders raising on a demo and a Notion doc.

But there's a second conversation happening, and it doesn't make the press release. Talk to enough general partners off the record and a pattern emerges: a real slice of experienced investors are passing on the category, deliberately and repeatedly. Not because they hate AI, most of them backed the foundation-model wave and made money doing it, but because the unit economics, the defensibility, and the durability of agent companies don't survive their underwriting.

This isn't the cartoon "old VC doesn't get it" story. The people sitting out tend to be the ones who've lived through a hype cycle or two. They watched the same dynamics in mobile (2010-2012), in marketplaces, in the first wave of "AI" startups around 2016 that were mostly thin wrappers on someone else's API. They learned that being early to a real technology and being early to a good business are different things. Their caution is a position, not a reflex, and it's worth understanding in detail, because their reasons map almost exactly onto the questions every agent founder will eventually have to answer.

Reason One: The Moat Problem

The single most common objection is the one founders least want to hear: what stops a competitor, or your own customer, from rebuilding this in a weekend?

A large share of agent startups are, structurally, an orchestration layer wrapped around frontier models from OpenAI, Anthropic, or Google, plus some prompt engineering, a few tool integrations, and a clean UI. That's a product. It is not obviously a moat. When the underlying intelligence lives in a model you don't own and your competitor can call the same API, the question of durable advantage becomes urgent fast.

Skeptical investors press on three classic sources of defensibility and find most agent companies short on all three:

a16z's own writing on the emerging architecture of LLM applications is unusually candid that much of the value in the stack accrues to the model providers and the infrastructure layer, with the application layer fighting hardest for durable margin. Investors who internalized that point read most agent pitches as a bet that this team will be the exception. Some teams will be. The sitters-out simply don't believe they can pick those exceptions reliably enough to justify entry valuations.

Reason Two: Margins at the Mercy of the Model Labs

The second objection is colder and more arithmetic. Agent companies have a supplier that is also, increasingly, a competitor, and that supplier sets their cost of goods sold.

Every time an agent completes a task, it makes model calls, and those calls cost money. A complex autonomous workflow can chain dozens of inference steps, each consuming tokens. In a per-task or per-outcome pricing model, the startup eats that cost directly. If a task that bills the customer a dollar costs sixty cents in inference, the gross margin looks nothing like the 80-90% software investors are used to, it looks like a logistics business with a software interface.

Two things make seasoned investors twitchy here:

First, the COGS line is controlled by a third party. Model providers can, and do, change pricing. Cuts have generally helped, but the dependency itself is the problem: your margin structure is a decision made in someone else's pricing meeting. Investors who watched cloud-cost surprises gut SaaS margins know how this movie can go.

Second, the supplier is moving up the stack. OpenAI, Anthropic, and Google are all shipping agentic features, computer-use capabilities, and assistant products of their own. An application-layer agent company is, in some cases, building the very feature its model provider will ship natively next quarter. That's a brutal place to be, and it's a recurring theme in the debate over whether agents are durable businesses or features in disguise. When your margin is thin and your roadmap overlaps your supplier's, the risk-adjusted return gets ugly quickly.

This is precisely the dynamic explored in the cluster's work on the burn-rate problem and on valuation haircuts when model costs compress margins, agents are expensive to run, and that expense doesn't disappear at scale the way it does in pure software.

Reason Three: Revenue That May Not Be Revenue

Here's the objection that separates the truly careful investors from the merely cautious ones. A lot of agent revenue right now is experimental-budget revenue, and experimental budgets get cut first.

In 2024 and 2025, large enterprises stood up "AI innovation" budgets and sprayed them across dozens of pilots. Agent startups booked that spend as ARR. But a one-quarter pilot funded by an exploratory budget is not the same as a renewed, line-of-business, mission-critical contract. The revenue-quality question, is usage revenue durable?, is the single most important diligence item the sitters-out run, and it's where many hot startups fail the test.

The tell they look for: net revenue retention and the renewal cohort. If a startup's customers are signing twelve-month deals and expanding, that's durable. If they're running 90-day proofs of concept that quietly lapse, the topline is a mirage dressed as growth. McKinsey's research on enterprise AI adoption keeps surfacing the same gap, most organizations are experimenting widely but capturing value narrowly, which means a lot of the spend agent startups are booking sits on the fragile, experimental side of that line.

Per-outcome pricing complicates this further. It's a genuinely attractive model when the outcome is clearly attributable and the customer trusts the attribution. But it also makes revenue lumpier and harder to forecast, and it invites disputes about whether the agent actually delivered the outcome. Investors who've been burned by usage-based revenue that cratered in a downturn want to see that the usage is structural, not discretionary.

Reason Four: Valuations That Assume a Perfect Decade

Even investors who believe in a specific agent company will sometimes pass purely on price. This is the most underrated reason for sitting out, because it's not about the technology at all, it's about the entry multiple.

When a seed-stage agent company raises at a valuation that implies a confident path to hundreds of millions in durable, high-margin revenue, the investor is being asked to underwrite near-perfect execution. The math leaves no room for the moat to be thin, the margins to compress, the model provider to compete, or the experimental revenue to churn. Stack those risks together and the probability-weighted return stops clearing the bar a venture fund needs.

The contrarian move, which several quieter funds have adopted, is simply to wait. Let the category's froth flow into the hottest names, let valuations reset, and pick up the survivors with proven retention at a sane price after the first wave of down-rounds. That down-round risk in over-funded agent startups is not hypothetical; it's the base-rate outcome of every category that raised faster than its fundamentals matured. PitchBook's tracking of venture valuation trends through the AI cycle shows the familiar pattern of late-stage and seed valuations decoupling from revenue during a hype phase, and reconverging, painfully, afterward.

Reason Five: The Liability and Reliability Overhang

The last reason is less about economics and more about the nature of the product. Agents act. Software that merely informs has a low blast radius when it's wrong. Software that autonomously books, buys, sends, deletes, or commits has a much larger one.

This connects directly to the cluster's work on agent reliability and agent security. An agent that's 95% reliable sounds impressive until you multiply it across a multi-step workflow, where compounding error rates can drive end-to-end success well below what an enterprise will tolerate. And when an autonomous agent makes a costly mistake, wires money wrong, leaks data, takes a destructive action, who is liable? The legal and insurance frameworks are immature, and conservative investors worry that a single high-profile agent failure could trigger a category-wide trust collapse and a wave of enterprise caution.

That's a correlated risk across an entire portfolio of agent bets, which is exactly the kind of risk a diversified fund hates. It's not priced into most rounds today, and the investors sitting out are, in part, waiting for the market to price it.

What the Sitters-Out Are Doing Instead

Sitting out the application layer doesn't mean sitting out AI. The funds passing on agent apps are often the same ones writing checks into the picks-and-shovels layer: evaluation and observability tooling, agent security and identity, orchestration infrastructure, and the data/memory systems agents depend on. The logic is straightforward, in a gold rush, the durable margin frequently sits with whoever sells the infrastructure, not whoever pans the river.

Others are simply being patient at the application layer: building relationships with promising teams, taking small information-rights positions, and waiting for the retention data to come in before committing real capital. A few are concentrating on deeply vertical agents in regulated industries, legal, healthcare, financial back-office, where integration depth, compliance requirements, and proprietary workflow data create the moat that horizontal agents lack.

What Would Bring Them Off the Sidelines

The sitters-out are not permanent bears. Ask them what flips a pass into a term sheet and the answers are consistent:

Hit those marks and the same investor who passed twice will move fast. That's the thing to understand about the quiet abstention: it's not skepticism about agents as technology. It's discipline about agents as a business, and the gap between those two is where the next few years of GaaS returns will actually be decided.

Insights Most People Overlook

References

More in Market