🔗 Part 1 of 2 — Under-the-Radar European AI Infrastructure

The LLMOps Testing Stack Is the New Cloud — Why Europe's AI Evaluation Layer Is the Best Seed-Stage Bet Right Now

By Sevda Anefi · Feb 2026 · 12 min read
Every dollar spent deploying an AI agent will eventually require a dollar spent testing that agent. Europe — with its regulatory DNA, engineering depth, and capital efficiency — is quietly producing the companies that will own this layer.

I. The Problem Nobody Wants to Talk About

The AI industry has a dirty secret: most production AI systems are untested.

Not untested in the "we didn't run a benchmark" sense. Untested in the "we have no idea how this behaves when a customer in Frankfurt asks it to process a GDPR deletion request in French at 3 AM" sense.

We are living through the fastest infrastructure buildout since cloud computing, but with a critical gap: the evaluation, testing, and observability layer barely exists. And the companies rushing to deploy LLM-powered agents — in banking, healthcare, legal, insurance — are flying blind.

Where AI Infra VC Dollars Go Today (Estimated)

Only ~7% of AI infrastructure investment targets evaluation and testing. Yet ask any ML engineer what keeps them up at night, and it's not training costs — it's "how do I know this thing works in production?"

This is a classic market timing dislocation. The capital is flowing to the wrong layer.


II. Why Now: Three Converging Forces

1. The Agent Explosion Demands a Testing Explosion

We've moved from single-turn chatbots to multi-step, tool-using AI agents that take actions — booking flights, filing reports, moving money. Each action is a liability surface.

The Complexity Curve — From Chatbots to Agents
2023
Chatbot Era
1 prompt → 1 response
Simple input/output pairs. Testing was straightforward: does the answer make sense? Failure was embarrassing, not catastrophic.
2024
RAG Era
1 prompt → retrieval → response
Added retrieval pipelines. Now testing means evaluating: is the right context retrieved? Is the answer grounded? Two failure surfaces instead of one.
2025–26
Agent Era
1 prompt → N tools → N actions
Agents chain tool calls, make decisions, take real-world actions. Testing surface area is combinatorial. Each tool call multiplies failure paths.
Result
Exponential Testing Surface
Complexity grows non-linearly
5 chained tool calls = hundreds of possible failure paths. This is why evaluation infrastructure is the fastest-growing segment of the AI stack.

When an agent chains 5 tool calls, the number of possible failure paths isn't 5 — it's combinatorial. The testing surface area of agentic AI scales exponentially, not linearly.

2. The EU AI Act Creates a Compliance Moat

The EU AI Act, enforceable from August 2025, isn't just regulation — it's a market-creation event for testing infrastructure. High-risk AI systems (credit scoring, hiring, medical devices) now legally require documented evaluation procedures, continuous monitoring and logging, bias and robustness testing, and human oversight mechanisms.

This is not optional. Compliance is now a line item in every enterprise AI budget. And European startups building testing tools have a structural advantage: they understand the regulatory language natively, they're building with compliance as a feature rather than an afterthought, and they're positioned at the doorstep of the world's largest regulated market.

3. The Evaluation Gap Is a Funding Gap

Most LLMOps/evaluation companies are dramatically underfunded relative to the problem size. Our signal analysis of eight European AI startups in the Jan–Feb 2026 window revealed something striking:

S-Curve Signal Score — Evaluation Stack (Jan–Feb 2026)
LangWatch
78
Agenta
75
Rhesis AI
66

These three companies are shipping at a cadence that rivals Series B teams, on pre-seed and seed capital. The signal-to-funding ratio is off the charts.


III. The Portfolio: Three Bets on the Evaluation Stack

#1
🔍
LangWatch
The "Datadog for AI Agents"
📍 Amsterdam
🌱 Pre-seed · €1M
📅 Feb 2025
Signal Score
78/100

LangWatch is building what I believe will become the default evaluation and observability platform for AI agent teams. Think of it as Datadog meets Cypress, but for LLMs — a unified platform where engineers and domain experts define evaluations, run experiments, test agent behavior, and monitor production.

Why this company, why now: In a single 30-day window, they pushed a core platform release with eval performance improvements, expanded guardrail setup, and shipped a versioned "Better Agents" CLI tool. The founders (Rogerio Chaves + Manouk Draisma) bring 25+ years combined from Booking.com and Lightspeed.

LangWatch Developer Loop
1
Developer writes agent
2
LangWatch SDK instruments traces
3
Define evaluations — correctness, safety, cost
4
Run experiments — A/B test prompts & models
5
Deploy to production
6
Monitor traces — latency, cost, failures
Guardrails trigger → feedback loop to step 3
Investment thesis: LangWatch is positioned at the intersection of the two fastest-growing budget lines in enterprise AI — evaluation and safety. With EU AI Act enforcement driving mandatory testing and the agent explosion multiplying the testing surface, LangWatch is building for a market that is about to 10x.
#2
⚙️
Agenta
The "Workflow OS" for LLM Development
📍 Berlin
🌱 Seed
📅 Jul 2023
Signal Score
75/100

Agenta is building the end-to-end development environment — from prompting through evaluation to production observability, with both cloud and self-hosted paths. Instead of optimizing one slice of the workflow, it's building the workflow OS.

Why this company, why now: Continuous merges, explicit versioning (v0.85.4 on Feb 12, 2026), plus GitHub Discussions about US region support and Enterprise SSO — classic indicators of a company preparing to move upmarket. These features don't get built for fun; they get built because enterprise customers are asking.

Agenta Platform Architecture
📝
Prompt Mgmt
🧪
Playground
📊
Evaluation
🔭
Observability
🚀
Production
☁️ Cloud 🏠 Self-Hosted
Investment thesis: Agenta is the rare LLMOps company building for workflow lock-in rather than feature competition. Its last disclosed round was July 2023 — nearly three years ago — which means either the team is extraordinarily capital-efficient, or they're about to raise. The signals strongly suggest the latter.
#3
🛡️
Rhesis AI
The "Trust Infrastructure" Play
📍 Potsdam
🌱 Pre-seed
📅 No public round
Signal Score
66/100

Rhesis AI isn't building another eval tool. It's building the testing infrastructure that turns compliance requirements into automated test suites — the missing bridge between "we need to test our AI" and "here are 500 test cases generated from your regulatory requirements."

Why this company, why now: Their Jan 2026 blog post describing a structured, multi-phase LLM testing cycle isn't content marketing — it's vocabulary creation. When you're the company defining how an industry talks about a problem, you're positioning to own the solution. No public funding rounds. Active GitHub development. This is the definition of a pre-seed opportunity with asymmetric upside.

Investment thesis: Rhesis AI is building the compliance-to-testing pipeline that every regulated AI deployment will need. In a world where the EU AI Act, DORA, and sector-specific regulations are creating mandatory testing budgets, the company that automates "requirements → test generation → regression detection" owns a critical workflow.

IV. The Structural European Advantage

Why are these companies European, and why does that matter?

Europe's Edge in AI Testing
🏛️
Regulatory DNA
EU AI Act native understanding · GDPR muscle memory · Compliance as feature, not afterthought
Engineering Depth
Strong CS university pipeline · Open-source culture (Berlin, Amsterdam) · Less demo culture, more shipping
💰
Capital Efficiency
Lower burn rates vs. SF · Longer runway per dollar · Forces product-market fit discipline
🏦
Market Proximity
Banks, insurers, auto OEMs · Regulated buyers with budget · Enterprise sales motion built-in

The US has the models. Europe has the testing, compliance, and reliability infrastructure. This isn't an accident — it's a structural outcome of building in a regulatory-first environment. And as AI regulation spreads globally, the companies that built for compliance first will have the global playbook.


V. The Trade

Portfolio Construction
Company Conviction Stage Why Mispriced
LangWatch 🟢 High Seed target Series A shipping cadence on €1M
Agenta 🟢 High Seed ext. / Series A 3 years since last round, continuous velocity
Rhesis AI 🟡 Med-High Pre-seed / Seed Category-forming + compliance wedge + unfunded

The bet: The evaluation and testing layer of the AI stack is about to undergo the same growth curve that observability (Datadog, New Relic) experienced in cloud infrastructure. The best time to invest in Datadog was 2012, when it was a small monitoring tool that most people ignored. The best time to invest in AI evaluation infrastructure is right now.


VI. What I'm Watching Next

LangWatch seed round timing — the shipping cadence screams Q2 2026
Agenta enterprise pipeline — SSO + US region = large ACV deals incoming
Rhesis AI first institutional round — the compliance-to-testing bridge is fundable today
EU AI Act enforcement actions — first penalties will catalyze panic-buying of testing tools
Incumbent response — when Datadog or Dynatrace launches an "AI eval" feature, it validates the category

💡 Bottom line: The AI evaluation stack is the most underfunded, highest-conviction opportunity in European AI infrastructure today. The companies building it are shipping faster than their funding would suggest, the regulatory tailwind is structural, and the market is about to wake up. The window to invest at seed prices won't last.