🧩🕸️✨ Agentic AI on the Web: Search, Action and the Mechanics of Reassembly

Team Members

Janna Joceli Omena, Richard Rogers, Giulia Tucci, Alice Figueiredo, Xiaohan Li, Yelyzaveta Terentieva, Maria Młocka, Xu Weiyu, Thomas Balster

Contents

Links

To Poster I AI Source Scraper I Graphics folder I Prompts-5concepts-AIsearch

Summary of Key Findings

1. Introduction

Agentic AI on the Web asks what changes when large language models stop only answering and start acting — searching, connecting, reasoning, and executing tasks across the web on our behalf. This project extends a longitudinal line of inquiry: a prior study found that GPT-4o exhibits homogeneity by default at the level of AI web search, repeatedly drawing from a narrow, recurring pool of sources unless sustained meta-prompting pressured it toward diversity (Omena, Tucci & Kishwar, 2026) — a finding that resonates with Jiang et al.'s (2025) much larger benchmark study of open-ended homogeneity across language models. That study took homogeneity as its object. This digital methods report takes it as its point of departure.

What has changed since is the emergence of a new answer layer above the web: retrieval, ranking, summarisation, citation and now action folded into a single interface. AI systems no longer only cite the web — they navigate it, connect to it, and act within it, on the user's behalf. Studying this shift means studying more than outputs. It means treating agentic AI as a condition for inquiry (Rodenbeck, 2026), a medium in which "output is not the answer, but the start of another question" — and tracing the operational chain behind that output: the prompts that shaped it, the sources it drew on, the permissions it required, and the labour it left with the researcher.

The project is organised around two subprojects:

• AI Source Vernaculars asks whether three AI platforms (ChatGPT, Claude, Gemini), given the same prompts across five media-studies concepts, search and cite the same web. (see https://www.digitalmethods.net/Dmi/SummerSchool2026AIVisualVernaculars)

• Application Mapping, Actions and Agentic Delegation asks what agentic AI is built to do — mapping the app ecosystems that equip it to act — and under what conditions it is actually allowed to search, connect, access, reason and act, documented through direct, situated engagement with three agentic AI systems. (see https://www.digitalmethods.net/Dmi/SummerSchool2026AIApplicationMappingActingDelegation)

The project addresses three critical epistemic moments in the use of agentic AI:

The mechanics of reassembly names moments where agentic AI search surprises us and expands thinking. Developed in response to David Berry's discussion of vibe coding and critical code studies (Centre for Digital Inquiry, University of Warwick, 2026), the term captures instances when AI search reassembles web traces while interrupting the prompted task — inviting the researcher to surrender to the process: a temporary suspension of the impulse to immediately take over, correct, measure, or instrumentalise the AI response.

Jagged intelligence names moments where agentic AI produces confident error — the uneven condition in which AI performs impressively on some tasks while failing unexpectedly on others, producing confident hallucination or plausible-sounding falsehood (Karpathy, 2025; Weeraman, 2026).

Agentic delegation names moments where agentic AI acts on one's behalf — not by retrieving or synthesising, but by performing tasks through embedded apps and platform services, making its infrastructural reach invisible in the process. Drawing on app studies, this third moment foregrounds how apps and platform interfaces redistribute agency through buttons, permissions, APIs and automated pathways (Gerlitz et al., 2019a, 2019b; Morris & Murray, 2018), extended here to the redistribution of action through agentic delegation.

2. Initial Data Sets

This project draws on three datasets, each built for a different point in the operational chain agentic AI is understood to move through. Project 1's corpus, comprises 266 source records captured across five media-studies concepts, three AI systems, and two search layers, gathered through repeated, same-settings prompting and a purpose-built AI Source Scraper, one folder per run; the records resolve to 219 unique URLs. Project 2's app-ecosystem dataset comprises 1,551 directory entries (1,540 unique apps) across ChatGPT, Claude and Gemini CLI, captured by hand through browser DevTools: each platform's directory tabs were reloaded with the Network panel open, the JSON response carrying the app list for that category (name, developer, description, icon) was copied out per tab, and the resulting per-category files were flattened into clean tables with a Python script. A third dataset has no equivalent directory to capture: Project 2's actions-and-delegation material consists of documentation generated through direct, situated engagement with three agentic AI systems (Claude Cowork, Gemini Flash Pro, Codex + Claude) across three case studies, yielding permissions maps, phased task plans, and action-and-negotiation logs rather than a scraped corpus.

3. Research Questions

4. Methodology

This project’s methodology focuses on investigating agentic AI as an emerging infrastructure of web navigation, knowledge mediation, delegation and action. It moves from asking how AI systems answer questions to examining the conditions under which they are able to search, connect, access, reason and act on our behalf. The method protocol traces the wider operational chain through which agentic AI systems are configured, prompted and allowed to operate.

The method protocol brings together three parallel and connected layers of inquiry: AI Search, App Ecosystem, and AI Actions and Delegation. It follows how search produces source vernaculars, how application ecosystems define the available field of action, and how delegated tasks unfold through settings, permissions, negotiations, interventions and residual user labour. At the centre, the mechanics of reassembly name the way AI platforms, prompts, permissions, sources, actions and human judgement are drawn together into apparently coherent outputs. The detailed procedures are unpacked on the subproject pages; AI Source Vernaculars and Agentic AI Application Mapping, Actions and Delegation. The protocol speaks for itself as a reproducible architecture for tracing AI-mediated search, action and knowledge reconfiguration.

5. Findings

· AI Source Vernaculars: shared lineage, unshared sources

The three AI search systems (ChatGPT, Claude and Gemini) produce three source vernaculars with almost no document overlap. There is no shared "ranking system", except for an agreement at the encyclopedia level: Wikipedia and Internet Policy Review concept pages, the only documents all three systems converge on, and only once agentic AI search (Layer 2) is active. Agentic AI search (Layer 2) expands into a more scholarly ecology — publishers, repositories, university domains — and recentres on canonical authors. This points to an old-web authority logic: AI search inherits Google's authority rule for who gets cited. At the level of names, authority still applies — the models converge on Gibson, Pariser, Helmond, Stiegler, exactly the canon a PageRank -style logic of accumulated reputation would predict. At the level of documents, however, the governing rule is not authority but availability. The canonical texts come from the freely crawlable surface web: course-page PDFs on lri.fr, self-archived copies on ResearchGate, the authors' own websites, doi.org redirects. Finally, robust source collection is not a default but an activation: with an underspecified prompt ("define the concept of technicity"), AI web search (Layer 1) surfaces a limited set of sources; it is only when agentic AI search (Layer 2) is activated — with the prompt specifying rules for how to search — that the volume and range of sources expand.

The diagram condenses the project's central finding into a single shape: across AI web search (Layer 1) and agentic AI search (Layer 2), the three models' flows converge into the same narrow set of canonical authors — the funnel — and then fan out into largely model-exclusive documents, with only six URLs shared between any two models (the black threads). The right side shows which sources each model drinks from. ChatGPT reaches the canon through publishers, journals and university domains; Gemini through repositories, preprints and reference works; Claude, strikingly, is divided across all five categories, with its largest single share coming from the authors' own web pages. Each model drinks from its own stretch of the web: same canon, different wells.

The domain geography behind the five source categories resembles what we have long known from search engines: retrieval bends towards location. The country-code domains are almost entirely Western European — the Netherlands leads (annehelmond.nl, uva.nl, uu.nl, tmgonline.nl), followed by the UK, France and Australia — with the United States present through generic .edu and .com infrastructure, and no domains from Latin America, Asia or Africa. This Dutch weighting mirrors the Amsterdam-school lineage of concepts such as platformization and the hyperlink economy, but it may equally reflect where the data was collected: the prompts were run from Amsterdam, and whether the retrieval geography follows the intellectual geography or the IP address is itself an open question. The reference-works category is thin but heavily drawn upon: just two domains, Internet Policy Review and Wikipedia, account for eleven citing records. And the residual "other web" turns out not to be residual at all — it is dominated by the authors' own websites (annehelmond.nl, jnd.org, danah.org, simondon.fr), including that of one of this project's authors (jannajoceliomena.net), who did not take part in data collection.

What differs across the three regimes is not how well each model searches, but what each believes searching is for: confirming the literatures (ChatGPT), verifying the canon it already holds (Claude), or supplementing the encyclopedia (Gemini). ChatGPT forewords its method and sweeps the fields — "Search Strategy — I checked: philosophy of technology/French theory, anthropology and archaeology, STS, media studies, software studies…" — nine literatures for technicity alone. Claude decides whether searching is needed at all, declining entirely for affordances ("'Affordances' is a term that doesn't require research") and elsewhere searching only to check itself: "I'll verify the key foundational sources for this concept rather than rely solely on memory." Gemini retrieves the periphery and anchors the canon from the weights: "I also explicitly consulted my parametric (training) knowledge for foundational historical authors (Mauss, Leroi-Gourhan…)." The epistemic regime, in other words, governs the source vernacular — style and retrieval are one operation.