Markdown source

Agentic Web

Agentic Web is the conference theme around making the public web usable by AI agents as an action surface, retrieval surface, interface layer, and data substrate. In this wiki it is a topic, not a standalone section: it connects scheduled talks, people, companies, videos, transcripts, slide evidence, and adjacent concepts such as agentic search, browser agents, MCP, sandboxes, catalogs, and agent-facing HTML.

A practical definition: the Agentic Web is what exists when an agent can discover a relevant web resource, understand the page or catalog well enough to choose an action, operate safely through the available interface, verify the result, and leave enough evidence for a human or system to audit what happened.

Why It Matters

The web was built primarily for people reading pages and clicking controls. Agents now need to search, compare, fill forms, inspect docs, use dashboards, purchase, schedule, and operate across services that may not expose clean APIs. That turns web pages, search engines, catalogs, HTML, screenshots, DOM trees, browser sandboxes, and identity controls into infrastructure for AI engineering.

For builders, this changes the product surface. A site is no longer only a human UX; it can become an agent-readable contract. A search index is no longer just a ranked list for people; it can become a planning substrate. A browser is no longer just a rendering engine; it becomes an execution environment with credentials, permissions, safety boundaries, and observability requirements.

How It Works

When To Use It

Use Agentic Web patterns when the work depends on public or semi-public web context, when the target service does not expose a sufficient API, when a human workflow must be automated through an existing UI, when agents need current facts beyond a private corpus, or when products want to be discoverable and usable by agents. Prefer direct APIs or MCP tools when the action is high-risk, repetitive, privileged, or available through a more stable machine interface.

Where It Is Useful

The pattern is most useful for research agents, shopping and catalog agents, browser-based operations, competitive intelligence, lead and company research, support workflows, data extraction, web-grounded coding assistants, documentation agents, and products that want AI systems to evaluate or recommend them. It is also relevant to agentic commerce, where identity, authority, and settlement become part of the web contract rather than an afterthought.

Origins In This Conference

The World's Fair schedule frames this as a multi-track theme rather than a single product category. Search talks emphasize finding the right live web context. Browser and computer-use talks emphasize operating existing websites. Catalog and HTML talks emphasize making web surfaces more legible to agents. MCP and sandbox talks add tool contracts and execution boundaries around those actions.

Related Scheduled Sessions

Search, Catalogs, And Web Data

Browser And Computer-Use Web

Agent-Facing Interfaces And Web Substrates

Related Resources

Related Topics

Related People

Related Companies And Tools

Design Patterns

Risks And Failure Modes

Open Questions

Evidence Boundary

Official schedule data is canonical for talk titles, speakers, dates, tracks, rooms, and inclusion in the schedule anchor list. YouTube videos, transcripts, local Whisper output, OCR, and reconstructed slide crops are supporting evidence. Treat transcript and OCR-derived claims as reviewable evidence, and verify important slide claims against the embedded slide image or reconstructed crop when available.