Title

Team Members

• Al Banchaabouchi Yassine
• Liliana Bounegru (King’s College London)
• Tommaso Elli
• Alex Gekker (UvA)
• Jonathan W. Y. Gray (King’s College London)
• Anne Helmond (Utrecht University)
• Prabhnoor Kohli
• Katie MacKinon (University of Copenhagen)
• Natalia Sánchez Querubín (UvA)
• Fernando van der Vlist (UvA)
• Nuoyi Wang (UvA / IHS Vienna)
• Esther Weltevrede (UvA)

Contents

1. Introduction

The central theme of this project is based around the notion that artificial intelligence, along with its agents, components, architecture, and ecosystem should be studied not as neutral infrastructures, but as conscious design choices. Therefore, a historiographic account of AI is necessary for a comprehensive analysis of its assemblages. Important historical narratives around AI can be studied in depth once the materiality of their source environments are made accessible. We further analyse how these materials came to be historical records by examining archives and the ecology of devices that create them. This allows us to gain a deeper understanding of what is brought in and left out in the creation of AI histories.

The project examines AI from the lens of historiography and archiving, specifically looking at:
1) AI agents & their architecture

2) The case of HuggingFace

3) AI integrations in the Photoshop interface

2. Research Questions

1. Which platforms, infrastructures, and repositories constitute critical source environments for future AI historiography?
2. What types of historical narratives about AI become possible if these materials are preserved and made accessible?
3. What are the methodological limits and blind spots of the current archival record of AI development

3. Methodology and initial datasets

The subproject followed an exploratory, comparative, and archival methodology designed to study AI agent platforms not only as technical systems, but also as evolving socio-technical infrastructures.

We first built a corpus of relevant cases by distinguishing between corporate agent ecosystems, such as OpenAI, Anthropic, Google ADK and related agent services, and alternative ecosystems, especially OpenClaw and the broader Claw family, including Hermes and NemoClaw.

This distinction was methodologically important because it allowed us to compare two different modes of agentic AI development: centralized, platform-led architectures on one side, and open-source, community-maintained architectures on the other .

Our empirical work consisted of systematically collecting and documenting multiple categories of digital sources. These included official documentation, GitHub repositories, ecosystem pages, developer resources, archived webpages through the Internet Archive, community platforms such as Discord, as well as Marketplace directories including software documentation.

The collected material was progressively transformed into structured research datasets. We decomposed each platform into comparable architectural components, including agents, models, tools, memory systems, orchestration layers, workspaces, integrations and governance structures, and organisée these observations within a common analytical framework.

Finally, these datasets were used to produce a series of comparative visualisations, including hierarchical dendrograms of agent architectures, heatmaps comparing architectural components across platforms, and network maps showing ecosystem relationships. These visualisations helped us compare the organisation of corporate and alternative agent ecosystems and identify their main similarities and differences.

Overall, this workflow enabled us to transform scattered online sources into a structured representation of contemporary AI agent ecosystems, contributing to the broader goal of documenting the material history of agentic AI for future historical research.

4. Findings

Archival footprint mapping

Mapping the public web presence of 150 notable AI companies, apps, and platforms across stakeholder-facing “sides” - developers, business, researchers, partners, policy actors, model/infrastructure stakeholders - shows that web-archive coverage is uneven not just company-by-company but side-by-side within the same company. Some platform sections (developer documentation, partner directories) are well represented in the archive; others (policy and safety-facing sections) are less prevalent or absent.

Hugging Face and the archival record

Zooming into Hugging Face as a key site of AI model circulation, the subproject examined both which of the top 50 models are archived and, more importantly, who archives them and through which mechanisms. The findings show archiving of model pages depends on a distributed mix of actors and tools like the Internet Archive’s own crawls, individual users via Save Page Now, automated processes like GDELT and ArchiveBot, and platform- or community-driven saving. Some model repositories are repeatedly, deliberately captured; while others enter the archive only incidentally, swept in by unrelated collecting projects (COVID-19 collections, national domain harvests, election-cycle crawls). This makes visible a finding that a model-only view of AI history would miss entirely: the historical record of Hugging Face is constructed by a distributed, largely uncoordinated set of archival actors, not by the platform or by any single deliberate preservation effort.

Reconstructing agent architectures

Comparing corporate agent providers (OpenAI, Anthropic, Google, with Amazon and Microsoft as cloud-based reference points) against community/alternative providers (OpenClaw and the broader “claw family,” including Hermes and NemoClaw) shows that both sides decompose into a comparable set of architectural components: model, tool/action use, context, orchestration, runtime environment, and governance. What differs is organization as corporate platforms tend to centralize this stack inside closed platform environments, while OpenClaw distributes functionality across a loosely federated set of independently maintained projects. The material evolution across both sides tracks a documented shift from single-turn instruction to persistent, multi-step autonomous operation.

Claw family use cultures

OpenClaw ’s own trajectory supplies a distinct empirical finding about legitimation: the platform moved from being called “a virus” by Microsoft’s CEO to Microsoft building its Scout product on it within roughly three months, while separately accumulating over 21,000–42,000 publicly exposed vulnerable instances, multiple disclosed CVEs (ClawBleed, ClawJacked), and documented incidents of malicious skills distributed through its ClawHub marketplace. In parallel, OpenClaw underwent rapid, large-scale adoption in China (Tencent-organized public setup sessions, subsequent local government policy support, and later restrictions on state enterprises), producing a wave of vernacular culture around it (the “raising lobsters” framing) that circulated across Chinese social media and international tech press simultaneously with the security disclosures.

Feature historiography: Photoshop’s Generative Fill

Because Adobe’s own release-note history is incomplete before version 27.0 (October 2025), the subproject reconstructed a timeline using YouTube Shorts as a primary source, cross-referenced against unofficial version-history sites, GitHub, Wikipedia, and videohelp.com’s changelog archive. The reconstructed timeline shows Generative Fill descending from the older, deterministic Content-Aware Fill (2010) through four distinct phases: introduction (2023, Firefly-powered, prompt replaces sliders), quality improvement (2024, sharper outputs, generated variations), customization (2025, choice of Firefly version, then partner models like FLUX.2 Pro), and creative control (2026, natural-language instructional editing, multiple partner models including Gemini, on-device inference). Across this arc the user’s role visibly shifts from manual operator, to prompt-writer, to model-selector, to an orchestrator who describes an outcome for a system

5. Discussion

Across all three empirical cases, we realised that AI history cannot be written by studying models alone. What becomes historically legible is not the model or the feature in isolation, but the ecology of archival actors (Hugging Face), the material architecture as well as the governance logic of agent stacks (agents), and the interface and tutorial culture through which a feature gets normalized (Photoshop). This throughline runs across the three subprojects, which function as three independent demonstrations of it rather than three unrelated case studies.

The project extends Helmond & van der Vlist’s platform and app historiography (2019, 2021) and Weltevrede’s notion of research affordances (2016) from social media into the AI domain, following van der Vlist, Helmond & Ferrari’s (2024/2025) framing of AI as infrastructure, models, and applications. The Photoshop subproject additionally draws on Manovich’s (2011) software studies argument that media software is not a single stable entity but a genealogy of individually evolving commands and functions, and on Bolter & Gromala’s (2003) critique of the myth of transparency for the window/mirror device. Together these anchor AI historiography as a natural extension of an existing methodological lineage.

Based on our findings and preliminary research, we worked on two posters that highlight the material and cultural traces of AI historiography across our three subprojects. The first poster brings together the footprint mapping of AI companies, the archival record of Hugging Face, and the feature history of Photoshop’s Generative Fill, using dense network and flow diagrams where the evidence is archival metadata, and annotated screenshots where the evidence is tutorial video. The second poster focuses on AI agent architectures and cultures, using dendrograms to compare how corporate and community-based agents (OpenAI, OpenClaw) are structurally organised, a star-history chart to show OpenClaw ’s rapid growth relative to the wider “claw family,” and images of use culture, including the lobster-costume install parties, to show that vernacular and meme-driven material is as much a part of the historical record as the technical architecture.

Poster 1 - https://drive.google.com/file/d/1dRbnkqTKIVxIs5w4NLqqrjs1fjFsdP-C/view?usp=drivesdk

Poster 2 - https://drive.google.com/file/d/1IqUpdnqoLVYa3EoeNb6SVbge0ogPFZtM/view?usp=drivesdk

6. Conclusions

The main finding across our three sub projects is that AI historiography is currently produced by accident more often than by design. Hugging Face model pages are archived incidentally, swept up by unrelated collecting projects, rather than targeted as AI objects in their own right. Agent architectures evolve and rebrand faster than standard web-archiving methods can track. Photoshop’s own release notes are incomplete for the very feature this project reconstructs. In each case, the record that future historians will inherit is the product of scattered, uncoordinated attention. This is the strongest argument for AI historiography needs to be practiced now and why it should be given more attention

7. References

Helmond, A. and van der Vlist, F.N. (2019) ‘Social Media and Platform Historiography: Challenges and Opportunities’, TMG – Journal for Media History, 22(1), p. 6–34. Available at: https://doi.org/10.18146/tmg.434.

Helmond, A. and van der Vlist, F.N. (2021) Platform and app histories: Assessing source availability in web archives and app repositories. In D. Gomez, E. Demidova, J. Winters and T. Risse (Eds), The Past Web: Exploring Web Archives. Cham, Switzerland: Springer, pp. 203–214. DOI: 10.1007/978-3-030-63291-5_16.

van der Vlist, F.N. and Weltevrede, E. (eds) (2025, May 28) Appification in the Age of AI: Exploring AI App Cultures and Economies (ASI Sprint Report Series No. 2). App Studies Initiative (ASI). DOI: https://doi.org/10.17605/osf.io/hv34x. https://appstudies.org/research-output/publications/asi-sprint-report-series/

van der Vlist, F.N., Helmond, A. and Ferrari, F.L. (2025 [2024]) Big AI: Cloud infrastructure and the industrialisation of artificial intelligence. Big Data & Society, 11(1): 1–16. DOI: 10.1177/20539517241232630.

Weltevrede, E. (2016). Repurposing digital methods: The research affordances of platforms and engines. PhD dissertation, University of Amsterdam. https://dare.uva.nl/document/2/168511

Bounegru, L., Gray, J., Venturini, T. & Mauri, M. (Eds.) (2018). A Field Guide to “Fake News” and Other Information Disorders. Public Data Lab. https://fakenews.publicdatalab.org

Manovich, L. (2011) ‘Inside Photoshop’, Computational Culture, 1. Available at: http://computationalculture.net/inside-photoshop/