KNOWLEDGE DOSSIER · Agents & protocoles

The Agentic Web

An agent can sequence decisions and tool calls to reach a goal. This autonomy is always bounded by the tools it receives and the rules set by the host application and responsible user.

Edition: 12 November 2026 Extended dossier Scientific review pending
AT A GLANCE
TypeAgentic workflows
AudienceReaders, organizations, developers
01

From answering to acting

Unlike a chatbot that only generates an answer, a tool-equipped agent can find a page, query a database, prepare a file or trigger an operation. Its results should be inspectable after each meaningful step.

02

Delegate without unlimited authority

Permissions should follow least privilege. Reading a document does not justify permission to send email or pay. Sensitive operations require explicit, contextual approval.

03

Instructions hidden in data

A retrieved page can contain text telling an agent to ignore instructions. That text is untrusted data: the host must separate analyzed content from the authority that defines allowed actions.

04

Practical example

An agent helps plan travel: it compares routes and drafts a shortlist. It neither books nor purchases without verifying the price, recipient identity and user approval.

05

Memory, planning and action boundaries

An agent may split a goal into tasks, consult resources, select tools and reassess progress. Each step creates a risk of confusing human instructions with untrusted Web content. Separate internal planning, external observations and authorized operations. A readable execution log should explain why a tool was invoked and with which parameters.

06

Scenario: search followed by approval

Suppose an agent finds course resources and proposes messaging participants. Reading pages and preparing a draft may be authorized without sending anything. Sending requires showing recipients, exact text and consequences, then waiting for explicit approval. The tool’s success response should be checked against the actual state in the destination service.

HOW IT FITS TOGETHER

Core steps

  1. 01Define the task
  2. 02Restrict tools
  3. 03Review outputs
  4. 04Approve sensitive actions
CRITICAL THINKING

Three questions to ask

  • Who can access which tools and data?

  • Can hostile page content change permissions?

  • Can the run be safely stopped or undone?

PRIMARY SOURCES

Read the source documents

Versions, reuse terms and document status can change. Check them on the original sites.