This workshop teaches effective work in agentic environments that maintain state, plan, use tools, create artifacts, and complete multi-step assignments, using Claude Desktop as the lab.

The day covers environment, context, delegation, trajectory control, capability boundaries, and reuse, providing a practical framework for directing substantial work without over-specifying or losing control.

Participants apply each concept in Claude Desktop through concise instruction, live demonstration, and hands-on work on one evolving simulated technology project.

Objectives

• Understand the components of an agentic working environment and the role each plays in sustained work
• Construct project context that guides behavior and remains reliable as information changes
• Define assignments that can be executed independently while preserving clear decision boundaries
• Guides work in progress through plans, intermediate artifacts, and meaningful intervention points
• Align tools and permissions with authority, reversibility, and risk
• Turning a successful one-time process into a reusable agentic working method
• Transfer these principles to other agentic AI environments

Target Audience

Technology Project Managers, Product Managers, Program Managers, PMO professionals, Delivery Managers, and leaders of cross-functional initiatives in technology organizations who want to manage substantive work in agentic AI environments without developing agents themselves.

Prerequisites & Experience

• Basic prior experience using generative AI tools
• An active paid Claude plan with access to Claude Desktop and Cowork; access must be verified before the workshop
• A personal computer for each participant

Agenda

The Agentic Working Environment

• From a single interaction to an environment that maintains state and acts over time
• The components of agentic work: context, tools, permissions, artifacts, plans, and memory
• The changing role of the user: from formulating prompts to shaping and directing a working environment
• Choosing the appropriate environment according to the persistence, actions, and autonomy required by the assignment
• Explored through a comparison of the same project assignment in Claude Chat and Cowork

Context as Operational Infrastructure

• Context selection rather than context accumulation
• The distinct roles of persistent instructions, project material, current state, and memory
• Establishing source precedence, identifying outdated or conflicting information, and preventing silent assumptions
• Context as a mechanism that shapes behavior — not only the quality of an answer
• Explored through the construction of a Cowork Project from a simulated project folder and participant modification of its source hierarchy and instructions

Delegation and Decision Rights

• Defining the required outcome without prescribing the complete path to it
• Constraints, expected artifacts, explicit exclusions, and evidence of completion
• Decision rights: what the environment may infer, propose, execute, or must escalate
• The balance between over-specification and insufficient direction
• Explored by turning a project request into an executable assignment and testing how Claude interprets its boundaries

Trajectory Control

• A plan as an inspectable hypothesis rather than a fixed execution script
• Intermediate artifacts as control surfaces for understanding the direction of work
• Local correction and redirection without restarting the assignment
• Strategic intervention points compared with continuous approval
• Explored by running the same assignment with different intervention policies and redirecting Claude during execution

Capability, Authority and Risk

• The distinction between technical capability and authority to act
• Reading, writing, executing, and making an external commitment as different levels of impact
• Reversibility, blast radius, information sensitivity, and the cost of an incorrect action
• External content as untrusted input and the role of prompt-injection awareness
• Placing approval at changes in risk rather than at every individual tool call
• Explored through a project-change scenario in which participants adjust a permission boundary, inspect the action trail, and retain human ownership of high-impact decisions

Turning a Successful Task into a Reusable Capability

• The distinction between local project context and a working method that should persist across tasks
• Projects, Skills, connectors, and scheduled tasks as different layers of the working system
• Extracting the invariant method: required sources, output contract, checks, authority boundaries, escalation conditions, and completion evidence
• Separating reusable behavior from project-specific context
• Explored by modifying a prepared Skill, applying it to a new scenario, and tracing the behavioral change back to the modified rule