🧩🕸️✨ AI Application Mapping: Actions & Agentic Delegation

Team Members

Yelyzaveta Terentieva, Maria Młocka, Xu Weiyu, Alice Figueiredo and Janna Joceli Omena

Links

To the main project: 🧩🕸️✨ Agentic AI on the Web I Poster I Application mapping interactive visualisation I Graphics folder

Contents

Summary of Key Findings

1. Introduction

Agentic AI systems (LLMs merged with web search, connectors, and the capacity to act across applications) are a genuinely new object of study. They are not chatbots: where a chatbot answers, an agent configures, navigates, edits, clicks, and executes. Public discussion of these systems tends to be binary and reactive, oscillating between embrace and rejection. This subproject proposes an alternative: to sit down with the medium and understand how it functions, how it operates, and what it costs — in tokens, in privacy, in control, and in environmental impact. This orientation draws on Omena’s (2022) medium-technicity, making room for computational media as carriers of meaning and developing sensitivity to their technical operations through situated engagement.

Within the digital methods tradition, Rogers (2013) describes the aim of digital methods as learning from and repurposing the devices of the medium for social and cultural research, under the general prescript to "follow the medium". Building on this tradition, and drawing on the concept of technicity-of-the-mediums, we argue that a medium can only be repurposed for research once the researcher has developed familiarity with its technical operations (Omena, 2022). For agentic AI, that familiarity does not yet exist, and existing methods do not simply transfer, because the object itself (an agent acting across tools, files, interfaces, and accounts) has no direct precedent in prior web research. The subproject therefore treats the researcher's own engagement with agentic systems: permission by permission, negotiation by negotiation as the primary record, and converts that engagement into a documented, reproducible protocol.

The central irony the project encounters is, however, not new. Bainbridge (1983) observed that "the more advanced a control system is, the more crucial may be the contribution of the human operator." Forty years later, working with state-of-the-art agentic systems, this project found exactly that.

2. Initial Data Sets

The application-mapping dataset comprises 1,551 directory entries, resolving to 1,540 unique apps, captured across ChatGPT, Claude and Gemini CLI. For ChatGPT and Claude, we mapped the consumer-facing app and connector directories, their comparable interface-agent surfaces, rather than their dedicated coding tools (Codex and Claude Code). Gemini CLI, by contrast, is a command-line environment, and its extension directory was mapped as such. The resulting dataset contains 375 ChatGPT entries (364 unique), 118 Claude entries and 1,058 Gemini CLI entries, providing a platform-level snapshot of the applications, connectors and extensions through which agentic action was made available in July 2026.

The Actions and Delegation dataset differs in kind. There was no pre-existing directory or corpus to capture because the object of study was the process of delegation itself as it happened. The dataset was therefore produced rather than collected through three participant-led case studies involving infographic-making, LinkedIn work and tool-making with Claude Cowork, Gemini Flash Pro and Codex + Claude. Researchers documented their engagement through prompt and chat histories, permission captures, screenshots and screen recordings, phased task plans, token-budget decisions, outputs, breakdowns, repairs and residual labour. The resulting material therefore treats the researcher's own engagement, permission by permission and negotiation by negotiation, as the primary record from which agentic action and delegation could subsequently be reconstructed.

3. Research Questions

RQ2a: What do the app ecosystems of ChatGPT, Claude and Gemini CLI reveal about whom agentic AI is being built and equipped to act for, and which everyday tasks remain outside its reach?

RQ2b: Under what conditions are agentic AI systems allowed to search, connect, access, reason and act on our behalf?

4. Methodology

Agentic AI Application Mapping

The application mapping method examines the applications made available through their interfaces and asks what forms of connection, delegation and action these infrastructures materially enable. Platform directories were captured through the browser interface and its Network layer, locating the Fetch/XHR requests through which app metadata were delivered and exporting the underlying JSON whenever available. These records were then flattened into comparable datasets containing information such as application name, developer, description and icon, followed by icon retrieval, data cleaning, merging and cross-platform matching. Applications were categorised through an LLM-assisted but researcher-revised process before being analysed through icon-composition views and app-platform networks. In this way, the application directory becomes a platform-native object for studying for whom agentic AI is being built, which tasks are becoming delegable, and which forms of everyday action remain outside its infrastructural reach.

Surrendering to agentic AI as a method

The three-phased method for understanding AI actions and delegation emerged from a methodological posture we call surrendering to the process. Participants were invited to work intensively with agentic AI for approximately two and a half to three days without beginning from a tightly predetermined analytical framework or continually interrupting the system in order to correct, measure or force it into an expected workflow. Surrendering here does not mean surrendering judgement. It refers to a temporary suspension of the researcher's impulse to take over the process, allowing the agentic system enough room to search, connect, request permissions, reason, encounter limitations, propose actions and redistribute work between user and machine. During this period, participants concentrated on documenting what happened through prompt and chat histories, screen recordings and captures, outputs, source traces, tool or connector use, permissions, moments of correction and other fieldnotes. Only after this period of situated engagement did participants look back across their sessions and reconstruct what they and the AI had actually done.

We treated methodological development as part of the inquiry itself, approaching agentic AI as a medium with which we first needed to become familiar (Omena, 2022). This familiarisation involved what has elsewhere been described, in response to Berry’s account of vibe coding, as “surrendering to the process” (Centre for Digital Inquiry, 2026): a temporary suspension of the impulse to immediately take over, correct, measure or instrumentalise the AI interaction. Our procedure therefore remained open so that it could emerge from engagement rather than precede it.

Participant-led auto-documentation operationalised this surrender without abandoning critical judgement: researchers engaged the systems on their own terms while documenting where the interaction led, when intervention became necessary, and what forms of delegation emerged through use. The dual position is therefore a choice rather than a compromise: delegation is experienced from the inside, in the moments of deciding what to hand over, when to intervene, and what to accept, and these decisions leave few traces available to an external observer. Working from that position, our expectation concerned limits, and we approached it through three case studies, each attempting full automation of a different kind of task in order to establish where that automation broke down and what it required of us instead.

A three-phase method for getting familiar with agentic AI

First, the agentic setting established the conditions for action through goal definition, apps or connector configuration, system setup and permissions. Second, prompt-budgeting and design translated goals into executable tasks, breaking work into phases while negotiating model constraints and the resources made available to the system. Third, delegated actions with negotiation traced what happened once agency was provisionally handed over: token budgeting, autonomy recalibration, permission negotiations, corrections prompted by model limitations, renewed user labour and eventual outputs.

The method protocol thus shifts the methodological object away from the AI answer alone and towards the conditions, negotiations and redistributions through which agentic action becomes possible. Surrendering becomes methodological precisely because critical reconstruction follows temporary suspension: first allowing the process to unfold, then returning to it to critically identify its operational patterns.

Three case studies

Three researchers agreed on three cases of working with agentic AI, selected to span different categories of task a person might engage in with such systems, divided between personal and professional use. Drawing on an Application Mapping visualisation of the current agentic tool landscape, we then determined which tools to include in each case.

• Professional. Using Gemini's agentic capabilities to act on our behalf on LinkedIn as part of a targeted job search, writing a post or sending a message.

• Professional. Producing a visual artefact about agentic AI, an infographic built in Figma, using Claude Cowork with minimal prompting.

• Personal. Building an application that sends a weekly Telegram message about techno events in Amsterdam, developed using Claude Cowork within a fixed token budget.

The cases were not designed to be comparable in scope. Their variation in task type, tool, and degree of surrendered control was deliberate: it allowed the same question about the limits of delegation to be tested against differently shaped work.

Data generation

As the object of study is the process of delegation itself, the dataset was produced rather than collected. As each engagement unfolded, researchers kept field notes, screenshots, screen recordings, captures of permission dialogues, successive prompt versions, and output logs. Prompt versions and permission captures were treated as primary material rather than incidental record. The former document how intent was reformulated in response to system behaviour; the latter mark the points at which the system paused to request authorisation and the researcher had to decide whether to grant it.

Analysis

Individual documentation was then compared and discussed collectively. We began by taking stock of what we actually held, asking which of our field notes, screenshots, recordings, permission captures, and prompt versions carried information about the work of delegation and which were merely residue of it. From there the discussion turned to observation: across the three cases, what had the system been able to do on our behalf, and what had it not. Setting those against one another let us locate the points at which delegation failed, and, in each instance, what we had to do ourselves to move past the failure. We then compared these points across the three cases, asking whether the interventions they required occurred at a similar moment. On that basis we distinguish the stages at which the human retains full control from those at which control passes to the system, and iterates on that distinction until it is resolved into a shared protocol of three steps, set out below.

5. Findings

Agentic AI Application Mapping

The Application Mapping begins from a simple infrastructural question: before agentic AI can act on our behalf, what has been made available for it to act through? By treating app, connector and extension directories as platform-native objects, the mapping examines the emerging conditions of delegation across ChatGPT, Claude and Gemini CLI. In July 2026, we captured 1,551 directory entries, resolving to 1,540 unique apps, and analysed their distribution across platforms, categories and interaction paradigms. The icon composition makes the density and thematic distribution of these ecosystems visible, while the app-platform network shows their clustering, platform specificity and limited points of overlap.

At first sight, the most striking difference is scale. Gemini CLI accounts for 1,058 of the 1,551 entries, approximately 68% of the entire dataset, compared with 375 for ChatGPT and 118 for Claude. The ecosystem as a whole also remains strongly technical: 913 entries, or 59%, fall within Software & Development, rising to 94% of Gemini CLI extensions. ChatGPT and Claude were mapped through their consumer-facing app and connector directories, their interface-agent surfaces, whereas Gemini CLI is explicitly a command-line environment for code agents. Its developer orientation is therefore less surprising than its extraordinary volume, which shows where extension-building around agentic AI is currently most intense.

Application mapping interactive visualisation here

The more revealing contrast lies between ChatGPT and Claude. Both expose agentic capabilities through conversational interfaces, yet their app ecosystems point towards very different forms of delegation. ChatGPT shows the clearest diversification towards everyday action, with substantial representation in Management (26%), Lifestyle (21%), Productivity (12%) and Creativity (7%). Its directory therefore resembles an increasingly general-purpose app ecosystem, extending agentic delegation into activities such as organising work, travel, shopping, content production and other everyday tasks. Rather than demonstrating who ChatGPT 's users are, the map shows what kinds of action its infrastructure is being configured to make delegable.

Claude produces almost the inverse picture. Despite being mapped through a consumer-facing interface comparable to ChatGPT 's, 78% of its apps are AI & Programming tools, many associated with MCP (Model Context Protocol), a standard that lets Claude connect to external tools, apps and data sources, and developer integrations. Claude therefore occupies an unusual position: an interface agent whose extension ecology remains structurally aligned with code-oriented agentic AI. This is one of the sharper findings of the mapping because it shows that a conversational interface does not necessarily produce a consumer-oriented app ecosystem. The interface may look general-purpose while the infrastructure beneath it remains overwhelmingly technical.

The ecosystems also show remarkably little convergence. Only eight app identities occur across platforms: Figma, ClickUp, Context7, Conductor, MeetGeek, Vibe Prospecting, Readwise and mcp-server-kubernetes. All belong to technical, professional or business-oriented contexts; no clearly consumer-oriented app crosses the divide. Cross-platform overlap is small and concentrated around infrastructural and professional functions. Rather than one agentic app ecosystem gradually converging around common services, the map shows highly platform-specific environments developing alongside one another.

This makes visible what we call an agentic gap. Domains closely connected to everyday and public life remain comparatively sparse, including News (6 apps), Health & Fitness (5) and Education (31). The contrast matters. Agentic infrastructures are expanding rapidly where actions can already be modularised, connected and executed through software, particularly within developer and business workflows, while many other forms of everyday delegation remain weakly represented. The promise that AI agents will broadly manage ordinary life is therefore much more uneven than the language of general-purpose agency suggests.

The visual form of the dataset provides a final clue to this uneven development. Within Gemini CLI, several “app icons” are photographs of people rather than polished product logos. These are GitHub profile avatars inherited from developers' repositories. What initially appears as a visual anomaly turns out to be evidence of the ecosystem's mode of production: a large part of the code-agent landscape is emerging through informal, repository-based contributions by individual developers. ChatGPT 's more branded app-store appearance and Gemini CLI's repository vernacular are therefore not merely aesthetic differences.

The network shows the app's distribution across platforms and categories, highlighting platform specificity. ChatGPT 's node branches into dense, varied clusters, pink Lifestyle apps like Booking.com and trivago, gold Productivity tools like Slack and Zoom, cyan Creativity apps like Canva and Adobe Photoshop, recognizable consumer brands spread across categories. Claude's branches are almost entirely green, AI & Programming, with labels like @sap-ux/fiori-mcp-server and blockchain-query. There's no real Lifestyle presence, no travel or booking cluster. Where ChatGPT connects outward into daily life, Claude connects inward into a developer's toolkit.

The map should consequently be read as the available infrastructure of delegation: the applications AI platforms expose, the domains into which agentic action is extending, and the domains that remain comparatively absent. In this sense, agentic AI application mapping provides a first layer for understanding the conditions under which agentic AI can act on our behalf, before examining what happens when users actually grant access, permissions and autonomy.

Agentic engagement: delegation negotiated, not total

Documenting agentic engagement through the three-phase method shows that delegation is a negotiation rather than a handover. Autonomy is purchased with specification: the more precisely a task is designed in advance, the fewer tokens are consumed in execution, so judgement migrates upstream into the prompt. Consent is technically present but not legible, since approving an application is not the same as confirming each action performed under it, and a ceiling on system-write permissions keeps final execution with the user. Roughly two-thirds of a desired outcome is produced through AI-mediated action; the rest is manual testing, corrections, refusals, and back-and-forth. Agency is redistributed rather than replaced.

Step 1: Agentic setting

The user prepares the conditions for AI-mediated action by configuring connectors, plugins, applications, and accounts; mapping permission types, from read-only access to full input control; and translating a vague goal into an actionable task with prompts, context, and expected behaviour. What the agent can later see, touch, and do is decided here. Documentation outputs: permissions maps, executable task lists.

Step 2: Prompt budgeting and design

User and AI co-design the operational pathway under real conditions of limited tokens, uncertain tool access, and anticipated failure. Prompting becomes a budgeting practice: the task must be precise enough to reduce waste while leaving room for agentic execution. Observation categories include anticipatory prompt design, decomposition of the task into phases, token budgeting, and specification of autonomy. Documentation outputs: meta-prompts, phased task plans, token-budget decisions.

Step 3: Delegated actions, with negotiations

The AI acts across tools, files, interfaces, and accounts while the researcher documents the negotiation: permission escalations, the capture, decide, click, and wait action loop, breakdowns and repairs, and the residual labour that remains with the user. Documentation outputs: action-and-negotiation logs, breakdown and repair records, residual-labour records.

The protocol's value is twofold. It renders AI action observable through concrete behavioural categories and provides a reproducible research procedure, allowing the same three phases and documentation categories to be applied across different agentic systems and tasks.

6. Discussion

The findings point to a single pattern across all three case studies. The integration illusion (Case 1), the illegibility of consent and configuration (Case 2), and the remaining 30% of manual labour (Case 3) are all examples of the same phenomenon: when the system reaches its limits, work is shifted back to the user in ways that the interface does not make visible. Delegation does not remove work; it redistributes and partly hides it. Configuration, permission-mapping, prompt design, token budgeting, corrections, and residual labour are the price of "autonomy". Agentic AI does not remove the researcher's expertise; it relocates it into system setup, supervision, calibration, technical translation, and judgement.

This is where Bainbridge's (1983) irony of automation becomes particularly relevant. The more capable the agent, the more the human's contribution is concentrated in the activities that automation cannot absorb: anticipating failure, verifying outputs, and deciding what counts as good enough. The three-phase protocol developed in this study makes this shift visible and possible to document, rather than leaving it as an observation based only on individual experience.

The relocation of labour has an infrastructural counterpart. Where earlier forms of web research could work substantially with publicly accessible pages, traces and platform data, agentic AI increasingly operates through account-gated and private environments: files, accounts, tokens, connectors and permissions. Acting on the user's behalf frequently requires granting access to spaces that were previously outside the research encounter. Web search persists as one layer of this environment, while agentic action extends into infrastructures that are not publicly observable. What is exchanged is access to personal or otherwise restricted data and environments; what is bought is speed and delegated action on tasks users might otherwise perform themselves. The delegation is driven by everyday workload, not incapacity, which is precisely why it is seductive, and why the terms of the trade deserve scrutiny. Not condemnation, but a clear-eyed account of what is exchanged.

The critique is therefore not "automation bad". It is a rejection of the separated vision, the fantasy of full automation where the human steps away. Humans still ferry data back and forth, and delegation is tested within a token budget, through iterative prompt refinement, at a measurable cost in privacy and control. This is Simondon's invitation: to get familiar with machines at different levels rather than stand apart from them, and it is the same posture as Rieder's theoretical and empirical work on understanding algorithms. Familiarity, not separation, is the methodological stance.

The significance for digital methods follows from this. Rather than beginning with the question of how agentic AI might be repurposed for research, we propose a prior methodological question: what does it take to become sensitive to a medium whose technical operations themselves carry meaning? This extends the “follow the medium” approach (Rogers, 2013) by making room for agentic AI as a carrier of meaning and developing sensitivity to its technicity before determining what methodological uses might be made of it (Omena, 2022). For agentic AI, this technicity cannot be apprehended from a distance: permissions, token costs, consent and delegated actions become visible through use. Surrendering to agentic AI is proposed here as a method for this first encounter: allowing the medium sufficient room to act so that its technical conditions, limits and forms of agency can become empirically perceptible.

These observations therefore document a first methodological encounter with agentic AI rather than a test of whether automation succeeds or fails. What becomes visible through that encounter is a redistribution of labour, judgement, access and control. The costs involved are not entirely new: questions of who owns data and who benefits from its circulation have accompanied digital platforms for more than a decade. What agentic AI changes is where that bargain takes place.

7. Conclusions

Full automation was not achievable in this research sequence. What the project shows instead is that the conditions under which agentic AI is allowed to act on our behalf operate at several levels: platform security architectures set hard limits, token economies set practical ones, consent mechanisms authorise broadly while providing little confirmation of individual actions, and prompts shape many of the judgements that "autonomy" then executes. Delegation, in practice, is negotiated step by step. Taken together, the application mapping and participant-led case studies make visible two dimensions of this emerging medium: the infrastructures through which delegation is being made possible, and the situated negotiations through which delegated action actually takes place.

Methodologically, the subproject proposes surrendering to agentic AI as a way of developing sensitivity to this medium through use. The resulting three-phase protocol, agentic setting, prompt budgeting and design, and delegated action with negotiation, turns that encounter into a reproducible research procedure for documenting the technical conditions, negotiations and redistributions through which agentic action becomes possible.

8. References

Bainbridge, Lisanne. 1983. ‘Ironies of Automation’. Automatica 19 (6): 775–79. https://doi.org/10.1016/0005-1098(83)90046-8.

Centre for Digital Inquiry, University of Warwick. (2026, May 14). Vibe Coding Workshop [Video]. YouTube. https://www.youtube.com/watch?v=tJkPO7IcDBk

Rogers, Richard. 2013. Digital Methods. The MIT Press. https://doi.org/10.7551/mitpress/8718.001.0001.

Omena, J. J. (2022). "Technicity-of-the-mediums". In Elgar Encyclopedia of Technology and Politics. Cheltenham, UK: Edward Elgar Publishing. Retrieved Aug 13, 2026, from https://doi.org/10.4337/9781800374263.technicity.mediums