Roko's Basilisk
Posts
AI's Middle Layer Wins

AI's Middle Layer Wins

Plus: Unitree goes mecha, Baglino's new bet, and Anthropic's security mess hits finance.

Roko's Basilisk
May 14, 2026

Here’s what’s on our plate today:

🧪 Why the companies in the middle are becoming AI's most powerful players.
📰 Unitree's mecha moment, Tesla alum's heat pump bet, Anthropic's Mythos rattles banks.
🛠️ Three tools worth trying: Rogo, Braintrust, Vellum.
🗳️ Poll: Who wins as AI labs push into deployment?

Let’s dive in. No floaties needed…

In partnership with

AI Agents Are Reading Your Docs. Are You Ready?

Last month, 48% of visitors to documentation sites across Mintlify were AI agents, not humans.

Claude Code, Cursor, and other coding agents are becoming the actual customers reading your docs. And they read everything.

This changes what good documentation means. Humans skim and forgive gaps. Agents methodically check every endpoint, read every guide, and compare you against alternatives with zero fatigue.

Your docs aren't just helping users anymore. They're your product's first interview with the machines deciding whether to recommend you.

That means: clear schema markup so agents can parse your content, real benchmarks instead of marketing fluff, open endpoints agents can actually test, and honest comparisons that emphasize strengths without hype.

Mintlify powers documentation for over 20,000 companies, reaching 100M+ people every year. We just raised a $45M Series B led by @a16z and @SalesforceVC to build the knowledge layer for the agent era.

Make Your Docs Agent-Ready

_{*This is sponsored content}

The Laboratory

TL;DR

Capability isn’t the bottleneck: Over 80% of organizations see no bottom-line impact from gen AI; 95% of IT leaders cite integration, not models, as the barrier.
Middleware is the power layer: Companies like Rogo, Braintrust, and Enter thrive by wiring foundation models into regulated, data-heavy industries.
Model makers want the middle too: OpenAI’s $10B deployment venture and Anthropic’s $1.5B PE deal signal the labs are moving downstream into middleware territory.
Politics shape access as much as tech does: Anthropic’s Pentagon standoff shows deployment decisions are increasingly entangled with policy, security, and institutional trust.
The stakes are structural: The more powerful AI gets, the harder deployment becomes, making the connective layer more essential even as platforms try to absorb it. If middleware can’t hold, a few labs own the full stack.

Why the companies in the middle are becoming AI’s most powerful players

For the past three years, the artificial intelligence race has been framed as a contest of capability, with companies like OpenAI and Anthropic pushing their models to ever-higher levels of performance. However, this framing has deflected attention from a bigger problem, one that compounds as AI models continue to evolve.

As AI models become increasingly powerful, implementing them into business functions and the daily lives of users has become ever more difficult. This means that capability is no longer the primary constraint, and the problem lies in translating that raw intelligence into systems that can operate reliably within the constraints of real organizations.

On May 4, both OpenAI and Anthropic made moves that suggest a growing recognition of this shift. O penAI finalized “The Deployment Company,” a $10B joint venture backed by firms including TPG, Brookfield, and Bain Capital, alongside a broader group of investors. Within hours, Anthropic announced a $1.5B partnership with Blackstone, Hellman & Friedman, and Goldman Sachs. Both ventures follow the same basic idea: embed engineers directly within companies owned by private equity firms, rebuild their workflows around AI, and use those portfolios as a built-in customer base.

Reflection AI sits in AI’s emerging middle layer, where the challenge is less about building models and more about securing trusted access to sensitive, tightly controlled environments. Photo Credit: The Information.

The message from model makers is clear: they built the highway, and now they are paying billions to have someone build the on-ramps.

To understand why the industry has reached this point, one needs to look at surveys that reveal that while adopting AI is simple enough, making it work within organizations can be much more complex.

The gap nobody budgeted for

According to a McKinsey global survey, more than 80% of organizations saw no tangible impact on their bottom line from generative AI, even as usage spread across departments. MuleSoft’s 2025 connectivity benchmark put a finer point on why: 95% of IT leaders who participated in the survey said integration problems were the primary barrier, with the average organization running roughly 897 applications and only about 28% of them connected.

Gartner projected that 40% of enterprise applications would feature AI agents by the end of 2026, up from less than 5% in early 2025. So, while the appetite for AI products is high, what is lacking is a delivery mechanism that makes adoption not only easier but also worth the effort.

This gap is now being filled by a new class of companies that do not build models or the thin application layers. These are companies that have planted themselves in specific industries and built the connections between what a foundation model can do in theory and what a regulated, data-heavy organization needs it to do in practice.

The companies in the middle

What’s striking is how each of these companies is solving a very different version of the same underlying problem: take, for instance, Rogo, a company working in investment banking. The company closed a $160M funding round in late April, and more than 35k bankers at firms like Lazard, Jefferies, and Rothschild & Co now use it daily.

Its AI agent handles the kind of work that junior analysts have traditionally done at 2 AM: screening deals, drafting memos, pulling data from financial databases, and preparing materials for client meetings. For end users, investment banking institutions, the underlying language model could have come from any provider. As such, what Rogo owns is the layer that connects that model to a bank’s internal systems, proprietary deal data, and the specific way investment bankers actually work.

And this layer is extraordinarily hard to replicate.

Another company operating in this space is Braintrust, which focuses on what happens after AI systems are deployed. As companies like Notion, Stripe, and Cloudflare scale AI features, Braintrust helps monitor and debug unreliable model behavior in production. This role is becoming increasingly critical as enterprise AI adoption grows.

Similarly, Enter is built around Brazil’s unusually AI-friendly legal system, where millions of structured, repetitive lawsuits create ideal conditions for automation at scale. The company now processes hundreds of thousands of cases annually for clients, including Itaú Unibanco, Santander Brasil, and Nubank, underscoring the value of environments with low deployment friction.

By contrast, Reflection AI operates where access itself is the bottleneck, building AI systems for tightly controlled environments, including classified government networks.

And at the more speculative end, Recursive Intelligence is pursuing the idea of AI-designed chips, reflecting a broader industry shift in which the real constraint is increasingly not the model, but the infrastructure and environments in which it can operate.

These companies then represent the middle layer, one that is becoming increasingly important yet operates under constant threat, since they do not control the platforms on which their businesses depend.

The threat to the middle-layer

The biggest threat to companies operating in the middle layers comes from the very AI labs they depend on for the underlying intelligence. If OpenAI were to build its own investment-banking agent, or if Anthropic were to introduce native observability and monitoring tools, much of the middleware layer could come under pressure.

From that perspective, the recent private equity-backed deployment ventures can be read less as an endorsement of this emerging category and more as a signal of competitive intent. The model providers, having established the core technology, appear increasingly interested in owning the layers that sit on top of it as well.

The future of these middle-layer AI companies may ultimately resemble the enterprise software industry more than the AI labs themselves. Earlier generations of software giants, such as Salesforce, Oracle, and SAP, attempted to capture the lucrative services and implementation layers surrounding their products. Yet, much of that work ultimately flowed to independent consulting and integration firms. AI may follow a similar pattern, where the long-term value lies not only in the models but also in the companies that adapt, monitor, deploy, and operationalize them within complex organizations.

That possibility helps explain why investors are assigning enormous valuations to companies operating between foundation models and end users. But it also suggests that the sector could evolve into something structurally messier and more competitive than current narratives imply. Many of these firms are being valued on the assumption that AI adoption will create entirely new operational layers across industries, even though it remains unclear how much of that market foundation model providers will eventually try to absorb.

The sector may also become increasingly shaped by politics and institutional alignment rather than solely by technology. Anthropic’s reported clash with the Pentagon over military restrictions on its AI systems, and the Defense Department’s subsequent move to phase out some of its products, illustrated how quickly deployment decisions can become entangled with questions of national security, policy, and access. By contrast, companies such as OpenAI, Google, and xAI appear to have taken a more flexible stance in government settings. That context makes companies like Reflection AI particularly notable because, despite limited commercial scale, they have secured access to highly restricted defense environments.

For the emerging middleware layer, the future may depend as much on regulatory alignment, trust, and institutional access as on technical capability itself.

The paradox at the center

The AI industry is entering a more complex phase than the one that preceded it. For a while, progress was easy to measure: better models, stronger benchmarks, more impressive demos. Now, the challenge is less visible and far more practical, centered on how that capability is actually put to work. A new set of companies, many of which barely existed a couple of years ago, has begun building the connective layer that makes AI usable in specific environments, from Brazilian courtrooms to Wall Street deal teams, from sensitive defense systems to the design of the chips themselves.

The open question now is how long they can hold that ground. As companies like OpenAI and Anthropic move deeper into deployment, some of these newer layers may come under pressure. At the same time, there is a quiet paradox at play. The more powerful AI becomes, the harder it is to use well in the real world, which makes the job of connecting it to everyday systems more important, not less. That tension is still playing out, and it is likely to shape what comes next.

Quick Bits, No Fluff

Unitree's mecha moment: Chinese robotics company Unitree is showing off a wearable mecha suit, a quirky milestone in the slow march toward human-robot integration.
Tesla alum bets on heat pumps: Former Tesla executive Drew Baglino, CEO of Heron Power, has founded a heat pump startup, signaling that climate tech is still drawing top engineering talent.
Anthropic's Mythos rattles banks: Anthropic's Mythos security incident has US banks scrambling to plug cyber holes, a reminder that frontier AI risks are now systemic financial risks.

Launch fast. Design beautifully. Build your company's website on Framer.

Framer helps teams design, build, and launch their marketing sites lightning fast.

With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster.

Trusted by companies like Miro, Bilt, and Perplexity