Knowlee vs H Company / Runner H (2026): Orchestration OS vs Frontier Action Model

Quick verdict. H Company (hcompany.ai, Paris) builds frontier action models — AI that plans and autonomously executes multi-step tasks in digital environments. Knowlee is the orchestration OS that governs action models, including Runner H, under a jobs registry, audit trail, kanban operator surface, and cross-vertical brain. These are different layers of the stack. If your question is "which action model should we run?" H Company competes with other model builders. If your question is "how do we govern, schedule, audit, and compound what those action models produce?" that is Knowlee's domain.

What each platform actually is

H Company (hcompany.ai, founded Paris 2024, $220M seed) was founded by Charles Kantor (CEO) and Laurent Sifre (former DeepMind research director). Backers include Accel, Bpifrance, Eric Schmidt, Xavier Niel, Bernard Arnault, Amazon, and Samsung. The company's stated mission is to build "frontier action models" — AI systems that reason, plan, and autonomously execute multi-step tasks inside digital environments (web browsers, desktop applications, APIs). Their first product, Runner H, is positioned as an AI agent for business process automation, QA, and RPA-style web workflows. The bet is that running tasks in real digital environments — not just generating text — requires a purpose-built model class, distinct from language-only foundation models.

Knowlee is an agentic operating system: a platform that schedules, governs, and compounds what agent models produce. It is not a model. It is the runtime above the model: a jobs registry with cron scheduling, an audit trail streaming reasoning per run, governance metadata (risk level, data categories, human-oversight flags, approval owner) on every job, a kanban the operator sits in, and a Neo4j Brain that accumulates everything every agent learns across all jobs and verticals so each new run starts from a richer state. Knowlee can consume Runner H or any action model as one of many tools in the agent fleet. The model does the acting; Knowlee is what makes that acting governable, schedulable, and compounding.

Architecture difference: action model vs. orchestration layer

H Company and Knowlee operate at different layers of the agentic stack — which is why the more useful question is not "which one?" but "in what order?"

H Company / Runner H: the action model layer

Runner H is trained to execute tasks in digital environments. Where a language model generates a response, an action model drives a browser, clicks buttons, fills forms, reads page state, and handles multi-step decision trees in real UIs. The model carries the capability to act. H Company's infrastructure wraps that capability with a deployment surface — you give Runner H a task description and it executes.

The strength of this model is execution fidelity in complex digital environments — workflows that break when you try to automate them with brittle selectors and fixed scripts. The limitation, from an enterprise governance perspective, is that a powerful action model running without an orchestration layer has no inherent audit trail, no scheduling, no cross-run memory, no operator kanban, and no governance metadata that an AI Act audit can interrogate.

Knowlee: the orchestration layer above the model

Knowlee's architecture is organized around jobs — typed, scheduled units of agentic work, each with declared inputs, outputs, governance metadata, and a streaming execution log. A job can call any model or any MCP-accessible tool, including Runner H. The jobs registry is the system of record for what the AI fleet is authorized to do, when, and under what oversight conditions. The Brain (Neo4j) captures what each job learns, so the next job starts from cumulative intelligence rather than zero.

This means Knowlee adds to Runner H — and to any other action model — the governance scaffold that enterprise deployment requires: who approved this job, what data categories does it touch, does it require human review before acting, what did it produce last time, and what did it learn? None of those questions are answered by the action model itself; they are answered by the orchestration layer.

Side-by-side comparison

Where H Company / Runner H wins

H Company is the right choice when the question is specifically about the action model layer:

• Execution fidelity in complex digital environments. If the workflow involves navigating real UIs — reading dynamic page state, handling CAPTCHAs, interacting with legacy web apps that have no API — a frontier action model trained for that task outperforms generic LLM tool calls.
• RPA replacement. Teams migrating off Selenium or UiPath-style bots to AI-native automation benefit from an action model that generalizes across UI changes. Runner H's training specifically targets this.
• R&D into action model capabilities. Research teams evaluating the frontier of what AI can execute autonomously — not just plan — benefit from direct access to a purpose-built model.
• Speed of task execution. When the bottleneck is pure execution speed in a digital environment, optimizing at the model layer (H Company) beats optimizing at the orchestration layer (Knowlee).

The honest caveat: a powerful action model without an orchestration layer is a capable but ungoverned executor. In regulated industries or multi-team deployments, that gap creates compliance and oversight risk.

Where Knowlee wins

Knowlee is the right choice when the question is about governing a fleet of agentic work across the organization:

• Enterprise governance from day one. Every Knowlee job carries declared risk classification, data categories, human-oversight requirements, and approval metadata. AI Act auditors get a native data model, not a log-parsing exercise.
• Scheduling and reliability. Cron-scheduled jobs, retry semantics, lock acquisition, and execution history are built into the runtime. Runner H has no equivalent — you build it yourself.
• Cross-run and cross-vertical memory. The Neo4j Brain means each campaign learns from the last. An account researched by one job enriches the next job's starting state. H Company has no compound memory layer.
• Operator visibility. The kanban surface shows what every agent is doing, what is waiting for review, and what was completed. Without an orchestration layer, operators build their own dashboards.
• Multi-vertical compounding. Sales signals inform talent sourcing; client delivery patterns inform outbound targeting. Knowlee routes all of that through one shared brain. Runner H operates on the task in front of it.
• Model-agnostic fleet. Knowlee governs Claude, GPT-4o, Gemini, Runner H, or any MCP-accessible model under the same jobs registry. You are not locked to one model vendor.

Decision framework

You need H Company when your primary gap is execution capability in digital environments — automating workflows that require an AI that can actually drive a browser or desktop app with human-level robustness.

You need Knowlee when your primary gap is orchestration, governance, and compounding — governing what your AI fleet produces, scheduling when it runs, auditing what it decided, and making each run smarter than the last.

Most enterprise deployments need both: H Company at the action model layer, Knowlee at the orchestration layer governing it. The two are complementary, not competitive.

For more on orchestration architecture, see agentic operating system explained and multi-agent orchestration patterns. For EU governance context, see agentic workforce platforms comparison 2026.

Book a 20-minute deployment review | See the platform | Compare with CrewAI | Compare with LangGraph