The agentic future
of Control Hub.

Admins needed more than
answers from a chatbot.

Cisco AI Assistant could help administrators find and understand information, but the experience stopped at read-only guidance. It could not take action on an admin’s behalf.

That limitation mattered in Control Hub, where the difficult part of administration is often not finding an answer—it is understanding the right scope, making a change safely, and knowing what will be affected before committing it.

I led the experience strategy for what should come next: an agentic administration model that could support real administrative work while preserving enterprise trust, governance, and control.

The existing assistant was
read-only—and context was confusing.

The starting point was Cisco AI Assistant (CHAI): useful for questions and discovery, but not designed to complete administrative actions.

CHAI also asked admins to explicitly add context before asking a question. In testing, that interaction performed very poorly. Categories such as configurations, data, reports, and workspaces overlapped conceptually, forcing users to understand our information model before they could ask for help.

That surfaced a more fundamental experience problem: the assistant was asking the admin to do the work of deciding what context mattered. For a future agentic experience, context needed to become something the system could understand and communicate—not another configuration task for the user.

Design the workflow around
the admin—not the AI.

That insight shaped the strategy. Instead of treating AI as a destination or asking admins to manually assemble context, I defined a model in which AI supports the administrative task across the surfaces where work already happens.

My current playbook defines two surfaces: embedded AI inside Control Hub and extended capabilities in external agents. The hybrid model describes the handoff between them, with a shared approach to authority and oversight.

What makes an agent
safe to adopt?

Unlike conventional software, agentic applications can evolve, gain capabilities, access tools, and act on behalf of users. We researched enterprise administrators, collaboration engineers, and developers to understand the mental models and governance needs this introduced.

I synthesized research findings and customer feedback into governance strategy, UX concepts, and administrative workflows.

Turning the strategy into
a practical design playbook.

I created an internal Agentic UX playbook to make the strategy usable by design, product, engineering, and platform teams.

I authored the strategy, audience framing, experience specifications for PM-defined use cases, audit and traceability model, and reusable skill packaging, and created the site that brings them together. The shared guidelines were authored by a colleague; I incorporated them into the playbook with that contribution distinguished from my own.

The shared workflow moves from understanding the goal and scope, to deciding on a plan, acting with the right authority, and returning a result the admin can verify.

Defining what external agents can do

I organized extended capabilities into five levels: public information, read-only admin context, troubleshooting, recommendations and preparation, and reserved execution. External experiences can support understanding and preparation. Changes such as restart, policy edits, access changes, remediation, and rollback remain on a Cisco-controlled surface.

Before action, Control Hub rechecks identity, scope, permissions, and current state. The experience keeps edits, approvals, denials, and outcomes traceable.

Keeping administrative actions
traceable across surfaces.

I defined a shared audit model that captures who requested, approved, and executed an action; what changed; why it was authorized; and the outcome, including rollback state.

Control Hub owns the authoritative record. The embedded experience provides the full log for investigation. An external agent provides a concise action receipt with status, a shared trace ID, and a link back to the full record.

The model distinguishes reads from writes. Read-only responses return provenance, scope, and freshness. Recommendations carry the rationale and trace ID into the next step. Edits and remediation create the formal record in Control Hub.

Making the playbook available
throughout the design process.

I packaged the playbook as a portable skill so teams can bring the strategy into AI-assisted design and engineering workflows. The site presents it as a downloadable SKILL.md containing the shared model, strategy, behaviors, use cases, references, and agent guidelines.

It is intended to support alignment before design, critique during development, and checks of handoff, recovery, traceability, and the result before launch. The packaging is my work; the included shared guidelines retain my colleague’s authorship.

This is a resource I’m developing and working with teams to adopt. As we begin supporting edit capabilities, Yankun Wang is reviewing audit logging. Adoption and audit design are ongoing; their effect on delivery has not yet been measured.