Palette Agentic

Kubernetes is complex. Troubleshooting shouldn’t start from scratch.

An in-chat AI agent concept for Palette, Spectro Cloud’s Kubernetes management platform, exploring how users could configure clusters, troubleshoot deployment failures, and reuse configurations through natural language while staying in control.

COMPANY | Spectro Cloud

ROLE | UX Design Intern (first summer internship)

DURATION | 10 weeks

TOOLS | Figma, Figma Make, Claude, Cursor, Bolt, Lovable, ChatGPT

METHODS | Secondary research, AI ecosystem review, interaction design, AI-assisted prototyping

STATUS | Exploratory product concept. Not shipped.

Making a complex platform easier to work with

Palette is used by developers, platform engineers, and IT operations teams to manage Kubernetes environments. Its workflows span cluster configuration, deployment, infrastructure management, and ongoing operations. Getting things done often takes deep technical knowledge and a lot of manual navigation, and when something fails, users have to interpret the error and decide what to do next.

The opportunity was to let users interact with these capabilities more naturally. An AI agent could take requests in plain language and help with three recurring tasks: getting started, recovering from failures, and reducing repetitive configuration.

Design challenge: Make complex workflows easier to navigate without hiding system behavior or taking decisions away from users.

The cluster-management workflow, with an AI-guided recovery path when a deployment fails.

Understanding agents, the product, and the people using both

My research covered three areas:

The AI agent landscape: What current AI capabilities and agent tools could realistically do.
Palette’s existing experience: How an agent could fit into the product through an in-chat assistant, rather than replacing the interface users already rely on.
Human-AI interaction guidelines: Google’s People + AI Guidebook and Microsoft’s Guidelines for Human-AI Interaction both stress user control, transparency, error recovery, and matching automation to the stakes of a task.

This was secondary research and product exploration, not customer interviews. It shaped how the agent should behave, not just how it should look.

Four principles for keeping users in control

I turned that research into four principles that gave configuration, troubleshooting, and recovery one consistent interaction model.

Guide: Break complex tasks into manageable steps.

Explain: Make recommendations understandable, not just actionable.

Confirm: Let users review consequential changes before proceeding. In infrastructure management, actions carry different levels of risk, so review should match the stakes.

Recover: Provide a clear next step when an error or recommendation doesn’t resolve the problem.

I turned that research into four principles that gave configuration, troubleshooting, and recovery one consistent interaction model.

Guide: Break complex tasks into manageable steps.

Explain: Make recommendations understandable, not just actionable.

Confirm: Let users review consequential changes before proceeding. In infrastructure management, actions carry different levels of risk, so review should match the stakes.

Recover: Provide a clear next step when an error or recommendation doesn’t resolve the problem.

Guide, Explain, Confirm, and Recover applied across a single request, from intent to outcome.

Three moments where AI could help

Rather than designing a general-purpose chatbot, I focused the in-chat assistant on three specific points in the workflow, each applying the principles differently.

01. Guided onboarding

Guide · Explain

Conversation breaks setup into manageable steps and explains recommendations alongside the existing configuration interface.

Guide · Explain

Conversation breaks setup into manageable steps and explains recommendations alongside the existing configuration interface.

The configuration stays visible while the agent explains each recommendation beside it.

02. Intelligent troubleshooting

Explain · Confirm · Recover

The agent translates deployment errors into clearer explanations and suggests recovery steps for users to review before anything changes.

Explain · Confirm · Recover

The agent translates deployment errors into clearer explanations and suggests recovery steps for users to review before anything changes.

An error, its plain-language explanation, and a proposed recovery step the user reviews before it’s applied.

03. Prefilled cluster setup

Confirm

A populated configuration gives users a starting point to inspect and customize instead of working through every setting from scratch.

Confirm

A populated configuration gives users a starting point to inspect and customize instead of working through every setting from scratch.

Prefilled settings stay fully editable, so users review a starting point rather than accept a black box.

The reviewed configuration, ready to deploy, with the agent still available for questions.

Designing for when the agent gets it wrong

I designed the interface in Figma and used AI-assisted prototyping tools to explore how conversations could unfold beyond a single response: follow-up questions, alternative recovery paths, and rejected suggestions.

Rejection is one of the most important cases to design for. An in-chat agent will sometimes misread what the user wants or propose something they don’t want. How it responds decides whether users trust it. One interaction I explored:

User: “Help me find a better way to configure this environment.”

Agent: Proposes a configuration approach.

User: “That’s not what I’m looking for.”

Agent: “Understood. Are you looking to simplify the configuration, improve performance, or address a specific issue?”

Instead of repeating the suggestion or going ahead anyway, the agent acknowledges the feedback and helps the user clarify what they meant. The user then decides whether to accept a revised suggestion or continue manually.

The rejection flow followed five rules:

Make suggestions easy to dismiss.
Let users correct misunderstandings.
Offer alternatives rather than forcing one AI-generated solution.
Keep users in control of consequential actions.
Never treat a rejection as permission to automatically try another action.

These were explored interaction patterns and design directions, not an implemented or tested feature.

AI-assisted prototypes used to explore how conversations unfold beyond a single response.

What’s next

I would evaluate the prototypes with both newer and experienced Kubernetes users around three questions:

Can users understand the recommendations? Do explanations help them interpret settings and troubleshooting steps?
Do users remain in control? Can they easily review, modify, reject, and correct suggestions?
Does the agent improve the workflow? How do task completion, time on task, and recovery success compare with the existing experience?

Before implementation, the agent’s need for system context and the reliability of its recommendations would also need to be evaluated against Palette’s technical constraints.

What I learned

Designing an AI agent for an existing technical product meant designing how it behaves, not just how the chat looks. Users need to understand why a recommendation makes sense, what it might change, and what happens when it’s wrong, including when they’re the ones saying no.

Guide, Explain, Confirm, and Recover gave the concept a consistent foundation for simplifying technical workflows without hiding the system underneath, or taking control away from the people running it.