VC & Startups/Nov 17 edition/Reporting & analysis
FundingFoundersEcosystems

FundingRounds, leads & terms

Remanence raises $30m seed co-led by EQT Ventures and FirstMark to train open models on AI agent traces

Paris startup Remanence, founded in September 2026 by three former H Company post-training leads, raised a $30m seed co-led by EQT Ventures and FirstMark to turn AI agents' execution traces into specialised open-weight models that customers own and continually retrain.

From the Sifted article of November 17, 2026 and cited sources: The $30m seed came about two months after the company was founded. [1] [2] [6] [8]

SOURCES[1] Sifted: EQT and FirstMark back trio of H Company alumni in… sifted.eu[2] Tech.eu: Synthesia and Datadog CEOs back AI startup… tech.eu[6] Sifted: Former H Company CTO seeks $30m for new startup sifted.eu

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
The $30m seed came about two months after the company was founded. [1] [2] [6] [8]
Original source
Sifted article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 17, 2026 article published
Evidence
13 sources across 10 sites

Reach: not measured yet

THE CORE IDEAS3 TAKEAWAYS
01

The $30m seed came about two months after the company was founded. EQT Ventures and FirstMark co-led it, with Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow also investing. Angel investors include Datadog CEO Olivier Pomel and both Synthesia co-founders. The final amount matches the $30m target Sifted reported in October. [1] [2] [6] [8]

02

Remanence plans to turn customers' agent logs into simulated enterprise training environments, then run reinforcement learning on them. Customers keep the model weights, and their data can stay on their own premises. The money will go to RL environments, evaluation tooling and a company-owned GPU cluster. [2] [3] [9]

03

The team's only public track record is H Company's open-weight Holo1 models for web-navigating agents, built on Qwen2.5-VL. Of the three founders, only Pierre-Louis Cedoz is publicly listed as a Holo1 author. Remanence has not released a product, a valuation or named customers, and cites only design partners in legal, customer support and enterprise software. [5] [7]

WHY IT MATTERS

investors made a large seed bet on post-training talent from a European agent lab. Implication: companies that run AI agents may get a way to own models built from their own logs, though performance and retraining costs are unproven.

Executive brief

Remanence is a Paris startup founded in September 2026, and on 17 November 2026 it announced a $30m seed round, a very large amount for a company that young. EQT Ventures and FirstMark co-led the round (Tech.eu). Its three founders previously led post-training of H Company's open Holo models. They want to turn the logs that AI agents produce while working ("execution traces") into specialised open-weight models that companies own and keep retraining (TFN). Angel investors include Datadog CEO Olivier Pomel and both Synthesia co-founders. The company has not disclosed a valuation, product benchmarks or named customers.

What changed and event timeline

  1. H raises a $220m seed

    The former DeepMind researchers behind H Company raised one of Europe's largest seed rounds. That company is where the Remanence founders later worked ().

  2. Holo1 released

    H published open-weight Holo1 models (3B and 7B parameters, built on Qwen2.5-VL) for agents that operate web interfaces. Pierre-Louis Cedoz is listed among the authors ().

  3. Founders leave H

    Cedoz left H at the start of September (). Remanence was registered as a Paris company (SAS) that month ().

  4. Raise reported in advance

    Sifted reported that the trio, including H's former CTO, was seeking $30m to post-train open-weight models ().

  5. $30m seed announced

    Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow also invested (;).

Capabilities and access

  • What it is: a platform meant to automate reinforcement learning (RL) and post-training of open-weight models using customers' agent execution traces.
  • Ownership: customers keep the model weights, and data can stay on their own premises (ActuIA).
  • Access: no public product, model name or version.
Read the full section
  • What it is: a platform meant to automate reinforcement learning (RL) and post-training of open-weight models using customers' agent execution traces. Models retrain as agents generate more data (TFN).
  • Ownership: customers keep the model weights, and data can stay on their own premises (ActuIA).
  • Access: no public product, model name or version. The company says it works with design partners in legal, customer support and enterprise software in Europe and the US (FrenchWeb).
  • Base models: none named.

Technical analysis for researchers and developers

  • Documented pipeline: agent traces are converted into simulated enterprise training environments, and RL is then run on those environments (TFN; Tech.eu).
  • Planned spending: RL environments, evaluation tooling and its own GPU cluster (ActuIA).
  • Reproducibility: there is no paper, model card or benchmark.
Read the full section
  • Documented pipeline: agent traces are converted into simulated enterprise training environments, and RL is then run on those environments (TFN; Tech.eu). How environments and reward signals are built from those logs is not described.
  • Planned spending: RL environments, evaluation tooling and its own GPU cluster (ActuIA).
  • Reproducibility: there is no paper, model card or benchmark. Holo1 is the team's only public track record: an agent split into a planning policy, a UI localiser and a validator, plus the WebClick benchmark (Hugging Face).

Claims and evidence

  • Round size and lead investors: announced by the company and repeated by Tech.eu, Axios Pro, TFN, FrenchWeb and ActuIA. No filing confirms them.
  • Valuation: not disclosed in any reviewed source.
  • Cheaper than proprietary models: a company framing with no cost data behind it (Tech.eu).
Read the full section
  • Round size and lead investors: announced by the company and repeated by Tech.eu, Axios Pro, TFN, FrenchWeb and ActuIA. No filing confirms them. Sifted's October report of a $30m target matches the final figure (Sifted). FrenchWeb's €25.5m is a currency conversion, not a separate figure.
  • Valuation: not disclosed in any reviewed source.
  • Cheaper than proprietary models: a company framing with no cost data behind it (Tech.eu).
  • Revenue and growth: the company has reported none.
  • Founder backgrounds: Holo1 authorship is publicly documented for Cedoz only (Hugging Face).

Context and prior work

  • H Company turmoil: three of H's five co-founders left soon after its 2024 round. A former Palantir executive later became CEO (Semafor). Co-founder Laurent Sifre stepped down as CTO (Sifted).
  • Crowded idea: learning from agent traces is a well-established idea in practitioner writing, for example LangChain's piece on continual learning for AI agents. It distinguishes learning at the model, harness and context layers.
Read the full section
  • H Company turmoil: three of H's five co-founders left soon after its 2024 round. A former Palantir executive later became CEO (Semafor). Co-founder Laurent Sifre stepped down as CTO (Sifted). Cedoz's later CTO role fits Sifted's "former CTO" label, though no reviewed source confirms the title directly.
  • Crowded idea: learning from agent traces is a well-established idea in practitioner writing, for example LangChain's piece on continual learning for AI agents. It distinguishes learning at the model, harness and context layers. Remanence works at the model-weights layer.

Limitations, safety and contested findings

  • Founder roles differ by outlet. FrenchWeb lists Hamza Benchekroun as president. To Morocco calls him CTO.
  • Angel lists also differ. Tech.eu omits Philippe Corrot (Mirakl) and Mehdi Ghissassi (AI71), whom TFN and FrenchWeb include.
  • None of the reviewed sources addresses the data-governance or safety risks of continuously retraining models on production agent logs.

Business and practitioner implications

  • For enterprises: the pitch is ownership.
  • Investor signals: a $30m seed shortly after founding shows investors paying for post-training talent from European agent labs.
  • What buyers should ask for: evidence that trace-trained open models beat frontier APIs on their own tasks, and the full cost of continuous retraining, including GPUs and evaluation.
Read the full section
  • For enterprises: the pitch is ownership. Firms that run agents keep the weights and turn their own logs into a defensible asset, instead of paying per call for closed models.
  • Investor signals: a $30m seed shortly after founding shows investors paying for post-training talent from European agent labs. The founder-operator angels (Datadog, Synthesia, Pigment, Mirakl) suggest enterprise-software distribution ties.
  • What buyers should ask for: evidence that trace-trained open models beat frontier APIs on their own tasks, and the full cost of continuous retraining, including GPUs and evaluation. Neither has been published.
  • Hiring: mainly in Paris, plus London and San Francisco (TFN).
FOLLOW THE EVIDENCE

The source trail.

Sources (13)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.