This session is not about a specific tool. It asks a more basic question: once agents enter everyday work, how should we reframe the work itself?

The answer I care about is structure: which objects deserve names, which steps can be compressed, and which mechanisms must keep human judgment intact. Without those boundaries, an agent becomes a more fluent but harder-to-manage black box.

Core claim.

The value of agents is not replacing every step. It is taking over frequent, describable, checkable intermediate work while helping people keep reusable representations, schemas, and mental models.

If a run ends without a structured artifact, and without making the next similar problem easier to judge, it was probably only a brief consumption of momentum.

01

Extract representations

Name context, roles, artifacts, and feedback instead of listing tool capabilities.
02

Form the schema

Compress each collaboration into Context, Role, Artifact, and Feedback.
03

Keep judgment

Let the agent lower load without taking over tradeoffs, verification, and responsibility.

Tool list

Models, plugins, commands, and buttons make the work feel scattered.

Work system

Objects, mechanisms, feedback, and artifacts make the work clearer over time.

An agent is not a tool list. It is a role inside a work system.

The same agent workflow can be read through four knowledge layers.

Audience.

  • —People using AI for writing, research, or development;
  • —Builders trying to place agents inside project delivery;
  • —Anyone tired of pushing every task into the same chat window.