Understanding Agentic AI: Search and Action Mechanics

Written by

JannaJoceliOmena

Aug 24—2026

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

New digital methods report on 🧩🕸️✨ Agentic AI on the Web: Search, Action and the Mechanics of Reassembly, 2026 Digital Methods Summer School and Data Sprint. This project brings together three parallel and connected layers of inquiry for practically interrogating agentic AI: AI Search, App Ecosystem, and AI Actions and Delegation. It reveals how AI 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.

Some key findings:

→ AI search takes different routes to the same destination, what we called AI Source Vernaculars. ChatGPT, Claude and Gemini AI search converge on who or what matters when it comes to media concepts (affordances; filter bubble; platformization; hyperlink economy; technicity), while diverging on the web sources through which that knowledge is retrieved and presented.

Claude is so smart that the model refuses to search the web to define affordances, even when explicitly told to do so. Any thoughts on that level of autonomy? 🤔

→ The mapping of Agentic AI apps reveals that ChatGPT offers a diverse range of apps focusing on everyday life, such as Management (26%), Lifestyle (21%), Productivity (12%) and Creativity (7%). 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. Agentic delegation seems to be developing into an uneven prospect: the promises of “do things for me” (ChatGPT) or “build things with me” (Claude and Gemini CLI) point toward developer automation.

The binary technological prospect of AI, good vs. threat, front-end vs. back-end usages, seems to keep its status growing ♻️

→ The three-phase method to engage with agentic AI, surrendering to the delegation process, shows instead 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.