Agentic Web
The Agentic Web is an emerging model of the Web in which people delegate goals to AI agents that can discover resources, interpret machine-readable capabilities, coordinate across services and perform authorized actions. The human remains the principal; agents become active intermediaries rather than merely returning documents or text. The label describes an ecosystem direction, not one protocol or completed technical standard.
Origin and context
Microsoft publicly framed an `open agentic web` at Build in May 2025. A later multi-institution survey organized the idea around intelligence, interaction and economics, while research on agent-facing interfaces argued that websites need explicit machine-readable actions instead of brittle visual navigation. W3C discussions subsequently examined semantics, identity, permissions, payments and browser enforcement. These sources converge on the problem but not on one architecture.
Why it matters
Today's Web primarily exposes interfaces for people and documents for crawlers. Agents pursuing multi-step goals need reliable discovery, scoped authority, durable task state, interoperable actions and evidence of what happened. Treating agents as another class of participant makes failures visible as infrastructure questions: who delegated an action, which service allowed it, what data was exposed, how consent is verified and how a disputed transaction can be traced or reversed.
Example
A travel agent could discover a carrier's structured booking capability, present options, obtain a user-approved spending mandate, authenticate with limited scope, reserve a ticket and return a signed receipt. The same goal attempted through screenshots and unrestricted stored credentials is still agentic browsing, but it lacks much of the interoperable, auditable infrastructure implied by the Agentic Web vision.
How it differs
AI browser / Agentic browser
An agentic browser lets an agent operate web interfaces from a user-agent surface. The Agentic Web is broader: it includes service-side capabilities, agent-to-agent interaction, identity, authorization, payments and governance across sites.
Agent2Agent Protocol
A2A is one communication protocol for agents. An Agentic Web may use A2A or other mechanisms alongside web semantics, tools and policy; the umbrella concept does not specify a single transport.
Agentic Commerce
Agentic commerce covers discovery, purchase and payment workflows. It is one high-impact domain within the wider Agentic Web, which also includes information, communication, productivity and public-service interactions.
Maturity and evidence
Maturity is rated 3. The term appears across a major platform announcement, independent research and W3C work, and concrete interface and protocol experiments exist. Yet W3C describes the field as early, definitions differ, protocols overlap, and no common conformance model spans discovery, delegated authority, identity, interaction, payments and accountability. The name is established; the proposed ecosystem is not mature.
Limits and open questions
Autonomous web action expands prompt-injection, privacy, fraud, access-control, traffic and liability risks. A declared identity does not show that an agent has current authority, and machine-readable actions do not guarantee correct intent or safe execution. Platforms may lawfully or technically restrict automation; policies vary by service and jurisdiction. The five-layer distributed legal infrastructure is one paper's proposal, not a governing standard. Evaluate individual protocols and deployments instead of treating the Agentic Web label as assurance.
Related terms
References
- Microsoft Build 2025: The age of AI agents and building the open agentic webMicrosoft · 2025-05-19 · class A
- Agentic Web: Weaving the Next Web with AI AgentsYang et al. / arXiv · 2025-07-28 · class B
- AI at TPAC 2025World Wide Web Consortium · 2025-12-09 · class A
- Build the web for agents, not agents for the webLù et al. / arXiv · 2025-06-12 · class B
- Distributed Legal Infrastructure for a Trustworthy Agentic WebChaffer et al. / arXiv · 2026-03-06 · class B
- The Agentic Web Requires New Normative InfrastructurePattison et al. / arXiv · 2026-06-09 · class B
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as ai agent design skill.
This term is also covered in the Skills Atlas as agentic planning task decomposition skill.
This term is also covered in the Skills Atlas as agent sandboxing skill.