Five‑Person Ooak Team Targets World’sLargest Workflow Library

The Library That Learns

Ooak Data, a five-person team from Paris, entered Y Combinator's Summer 2026 batch with a mandate to acquire 300 real enterprise workflow datasets in six months, the raw material for Alexandria, a library of digital twins that frontier labs now treat as a prerequisite for shipping reliable AI agents.

The company was founded in 2024 by Thomas Aubry, Grégoire Lamy, and Pierre-Louis Vouteau with a single thesis: the missing layer isn't more model capacity, it's the exhaust of real business operations, captured, anonymized, and turned into environments where agents can learn and be evaluated.

They named the project Alexandria, after the ancient library that tried to collect every scroll in the known world. Ptolemy's librarians seized books from ships, copied them, kept the originals, and returned the duplicates. Ooak's forward-deployed engineers plug into partner companies, extract the full multimodal workflow — documents, chat logs, project boards, org charts, tool calls — run it through an automated anonymization pipeline that preserves structure and relationships while stripping identities, and hand back a digital twin the company can safely share. The original stays with the owner. The copy joins the library.

The mission statement on their site is blunt: "The promise of AI agents is still unmet in complex, real-world business environments. Our mission is to make it possible for agents to reliably accomplish concrete and useful tasks." That reliability target is why they joined Y Combinator, not to chase a benchmark score, but to build the evaluation infrastructure that frontier labs already require for shipping agent products. Over the next six months, the team plans to acquire and anonymize more than 300 full data ecosystems from operating companies. Each ecosystem is a complete digital twin: every tool, every handoff, every exception path a human worker navigates. Most evaluation frameworks test single-turn Q&A. Ooak builds multi-step, multi-tool environments calibrated on the latest frontier models, designed to expose weaknesses rather than confirm strengths. As models improve, the environments evolve. You are always testing at the edge of capability.

The stack reflects that ambition. A document QA and editing platform will let hundreds of external contributors refine the anonymization output and uphold quality standards. An internal tool helps forward-deployed engineers extract partner data efficiently and safely. The anonymization pipeline itself is multimodal from the start — not text with modalities bolted on later — because real work isn't text-only. A procurement workflow lives across emails, PDFs, ERP systems, chat threads, and approval chains that shift with organizational changes. Synthetic benchmarks test what models can do in theory. This data tests what they do in practice.

The ancient library had Callimachus and his Pinakes, the first great catalog, separating information from the mechanism to retrieve it. Ooak's catalog is the RL environment layer: expert-level tasks that turn raw workflow into a signal a model can learn from. The research lab they plan to launch in early 2027 will sit on top of that catalog, turning the library into a service layer for companies that want to automate their own workflows with AI. First, though, the collection has to grow, because the benchmarks the industry relies on don't measure what makes agents reliable in production.

The Reliability Gap Is a Measurement Gap

Current agentic AI benchmarks evaluate task completion accuracy while overlooking the enterprise requirements that determine whether an agent survives deployment. A 2025 arXiv analysis of six leading agents across 300 enterprise tasks found a 37 percent performance gap between lab tests and production deployment. Single-run evaluation masks brittleness: GPT-4-based agents drop from 60 percent success at pass@1 to just 25 percent at pass@8, insufficient for enterprise deployment where reliability is paramount.

The pass@k metric, introduced in τ-bench (Yao et al., 2024), exposes this directly. The arXiv authors propose CLEAR (Cost, Latency, Efficacy, Assurance, Reliability), a framework that addresses these gaps. Reliability (R) assesses consistency through pass@k, with production deployment requiring pass@8 ≥ 80 percent for mission-critical applications. Expert evaluation (N=15) confirms CLEAR better predicts production success (correlation ρ=0.83) compared to accuracy-only evaluation (ρ=0.41).

Amazon's experience reinforces the point. Since 2025, thousands of agents have been built across Amazon organizations. AWS engineers found that traditional LLM evaluation methods treat agent systems as black boxes and evaluate only the final outcome, failing to provide sufficient insights to determine why AI agents fail or pinpoint root causes. Their response: create golden datasets for regression testing generated synthetically using LLMs from historical API invocation logs upon user queries. Simulation datasets contain user query and ground truth intent pairs collected from anonymized historical customer interactions. This is real-world workflow data — not synthetic benchmarks — driving evaluation.

IBM's production requirements make the causal chain explicit. Production systems must be reliable under load, observable in real time, governed across teams, secure across identity boundaries, compliant with policy and regulation, and measurable against business KPIs. Evaluation and optimization form two loops: an experimentation loop during build and a runtime optimization loop after deployment. This process is how organizations move from "it works in a demo" to "it works in the real world."

The technical literature identifies specific failure modes that workflow data addresses. Poorly defined tool schemas and imprecise semantic descriptions result in erroneous tool selection during agent runtime, leading to invocation of irrelevant APIs that expand the context window, increase inference latency, and escalate computational costs through redundant LLM calls. Amazon defined cross-organizational standards for tool schema and description formalization, creating a governance framework with mandatory compliance requirements. They also implemented an API self-onboarding system that uses LLMs to automate generation of standardized tool schemas and descriptions, significantly improving onboarding efficiency.

Containerization provides a consistent and reproducible runtime environment that eliminates configuration drift and ensures workflows behave identically in development, staging, and production. Separation of reasoning from tool execution ensures deterministic behavior, eliminates missed tool calls, and significantly improves stability over repeated runs. Complexity is one of the biggest threats to reliability and maintainability of agentic AI workflows; domain-specialized approaches outperform general-purpose architectures, achieving superior cost-normalized performance (260.4 vs. 14.5–58.0 CNA) and reliability (72.8 percent vs. 52.1–64.5 percent pass@8).

While 85 percent of companies experiment with generative AI, only a small fraction deploy agents in production, with most projects abandoned after proof-of-concept stages. The causal link is clear: benchmarks that measure only accuracy produce agents that optimize for the wrong objective. Real-world workflow datasets (capturing multi-tool sequences, exception paths, human decision points, and the messy context of actual business operations) enable evaluation on the dimensions that determine production viability: consistency under repetition, cost awareness, governance compliance, and graceful degradation. That is the data Alexandria is built to supply.

Seven Roles, One Summer

The mandate demands more hands than the founding trio can supply alone. Ooak Data lists seven open roles on Work at a Startup and its YC jobs page, a deliberate sprint to expand before the summer ends.

The most visible listing is a Marketing Lead, offered full-time or freelance, with a salary band of €50K–€75K and location flexibility that maps the company's geographic strategy: Paris, Île-de-France, San Francisco, and "Remote (US)." That spread is not accidental. Paris anchors the founding team and European data-privacy expertise (GDPR, nLPD compliance are baked into the anonymization pipeline). San Francisco places recruiters and engineers within walking distance of the frontier AI labs that consume Alexandria's datasets. The U.S. remote option widens the talent pool to time zones that can support 24-hour data-ingestion cycles without requiring relocation packages.

Aubry frames the push bluntly on the careers page: "This summer we need to build our elite squad to move extremely fast and confirm our leading position on company data." The language mirrors the operational reality: each new hire must onboard into a workflow that ingests raw enterprise logs, strips PII, structures multi-tool traces, and ships evaluation environments to model providers. There is no ramp-up quarter; the first 300 ecosystems are the ramp.

The remote-first model also solves a hiring economics problem. Competing for ML engineers and data-privacy specialists in Paris or the Bay Area means bidding against labs that pay far more. Ooak Data cannot match those bands. By hiring remote-first across the U.S. and Europe, it accesses senior talent who have already exited big-lab compensation structures but still want ownership-level impact, exactly the profile the company courts with lines like "No micromanagement, no unnecessary process. You will have real impact from day one."

Culture becomes the differentiator. The careers page lists operating principles: hire owners, admit uncertainty, prefer right-slow over wrong-fast. In practice, that translates to giving a data-pipeline engineer authority to own core technical decisions, or researchers freedom to publish findings from Alexandria. The trade-off is explicit: lower cash compensation than the labs, higher autonomy, and a direct line to the 300-ecosystem target that feeds the 2027 research-lab milestone.

The hiring surge is therefore not a headcount metric; it's a throughput requirement. Seven roles this summer, each plugged into a distributed ingestion pipeline, are the minimum viable team to turn 300 raw enterprise environments into the evaluation infrastructure that frontier labs already treat as a dependency.

When the Lab Opens in 2027

Ooak Data's founders have set a concrete milestone: launch an applied AI research lab in early 2027. The lab is the logical extension of Alexandria. The corpus of digital twins becomes the lab's raw material.

The plan, as stated on the company's Y Combinator profile, is to "enable new services that help companies around the world automate their workflows with AI." In practice, that means turning the evaluation environments Ooak already builds for frontier labs into products enterprise teams can consume directly. Today the company offers three tiers: RL environments grounded in real enterprise data for model providers, digital twins that let enterprise AI teams stress-test agents before production, and evaluation infrastructure for startups that lack data pipelines. The lab will productize all three.

Thomas Aubry, the CTO, wrote in the company's research blog: "80 percent of AI projects fail in production because evaluation conditions do not match deployment conditions." The lab's charter is to close that gap systematically. Its research agenda (evaluation methodology, dataset design, and the benchmark-to-reality delta) mirrors the problems Alexandria was built to solve.

This is not a pivot. The company's website describes its core mission as "the data infrastructure that frontier AI labs use to train and evaluate their agent models." The lab makes that infrastructure addressable to the buyers who ultimately pay for agent reliability: the enterprises deploying the models. If Alexandria is the library, the lab is the reading room where the books get stress-tested against the questions businesses actually ask.

The company's existing distributed posture suggests remote recruiting across Paris, San Francisco, and U.S. time zones. Ooak's anonymization pipeline is designed to satisfy GDPR, CCPA, and the Swiss nLPD without stripping the operational fidelity that makes the workflows useful.

Everyone Is Building the Same Bridge

The incumbents are not waiting. Scale AI and Labelbox — the two platforms that built the data layer for the current generation of foundation models — have each spent the last year retooling for agents. Scale claims 90 percent of leading generative AI model builders as customers. Labelbox says it partners with over 90 percent of leading U.S. AI labs. Both now position themselves as the evaluation and RL infrastructure for the next phase.

Scale launched Scale Labs, a dedicated research arm, and published SWE-Bench Pro to stress-test agentic coding. It expanded its data engine for physical AI (robotics and autonomous systems) and locked in enterprise partnerships with Mayo Clinic, Meta, the Center for AI Safety, and British Petroleum. The company's own marketing states that most AI deployments in enterprise and government fail, and that reliable AI has no shortcuts. Its hiring reflects the pivot: Scale AI's board data shows 10 roles added in the past week, including a Senior Staff Frontier Agents Engineer at $288K–$360K and a Tech Lead Manager for ML systems at $290K–$363K.

Labelbox moved faster on evaluation tooling. In Q3 2024 it shipped leaderboards for image, speech, and video generation models, an arena-style model comparison UI, built-in AI critics for automated feedback, and Labelbox Monitor for real-time labeling operations visibility. It expanded the Alignerr network (2.6 million knowledge experts, 25 percent with advanced degrees) and added on-demand human evaluation services inside the platform. In February 2026 it acquired Upcraft. By August 2025 it had released R-ConstraintBench to stress-test reasoning under interacting constraints and launched Labelbox Evaluation Studio for real-time model performance feedback. The platform now benchmarks 22 models across 47 reporting configurations with 820 expert-authored problems; roughly 20 percent of items remain above frontier ability.

Model providers are moving up the stack. OpenAI and Anthropic have expanded enterprise deployment arms, announced financial services partnerships, and shipped agent tooling targeting enterprise workflows. The strategic logic is clear: buying services firms would let them move from infrastructure vendors to full-stack enterprise players, controlling the relationship from model selection through implementation to workflow control. Both have partnered with asset managers to market enterprise products more aggressively. The hiring signal is loud: OpenAI added 34 roles in the past week, Anthropic 38, with research engineer bands reaching $500K–$850K at Anthropic, Zero G Talent's data shows, and $295K–$585K at OpenAI, Zero G Talent reported.

RPA veterans are pivoting in parallel. UiPath announced partnerships with Snowflake, Google Cloud, Nvidia, and OpenAI to integrate agents with enterprise data platforms, and a validated partnership with Databricks blending trusted data with automation. The three dominant RPA vendors — UiPath, Automation Anywhere, and SS&C Blue Prism — are all reinventing themselves as agentic AI platforms in a rapidly growing market.

Ptolemy's librarians didn't just collect scrolls; they built the catalog that let scholars find what they needed. That catalog serves as the RL environment layer. When the reading room opens in 2027, the question won't be whether the benchmarks have changed. It'll be whether the agents that train on Alexandria finally work when the approval chain changes on a Tuesday.

Working in frontier tech? Zero G Talent tracks the openings: see every open OpenAI role, browse frontier tech jobs, openings at Anthropic and Scale AI, and the people building the field.