AHAnli HuangAI Lab / Design Note← AI Lab

AI Lab · Design Note

Designing Lite Agents for Creative Workflows

What a TikTok Design AI Workshop taught me about building AI-native products.

During TikTok’s internal Design AI Workshop, our team explored how AI-native Agents and embedded Lite Agents could coexist across creator marketing workflows. This article summarizes the design principles, interaction patterns, and product thinking that emerged from the workshop.

CategoryAI Lab
Year2026
LocationTikTok Design AI Workshop
Reading time6 min read
Scenes from the TikTok Design AI Workshop, including talks, group discussions and concept presentations
TikTok Design AI Workshop · Cross-functional exploration and concept development

01 · Context

One ecosystem needs more than one form of AI.

TikTok One already contains AI capabilities across inspiration search, creator collaboration, creative production, campaign delivery, and reporting. Yet existing workflows still move unevenly between traditional interfaces and AI interactions.

The workshop started from a product problem rather than an event theme: different tasks require different levels of autonomy. A multi-step campaign goal may benefit from an Agent that owns the workflow, while an in-context rewrite or review should help immediately without taking users away from their work.

02 · The design question

How should AI-native Agents and embedded Lite Agents coexist?

Primary workflow

AI-native Agent

Owns complex, goal-driven work across multiple steps and modules.

SHARED CONTEXT ↔ SHARED OUTCOMES
Embedded assistance

Lite Agent

Accelerates one bounded decision inside the interface already in use.

03 · Insight 01

AI-native Agent ≠ Lite Agent.

The distinction is not about model intelligence. It is about the relationship between the AI, the task, and the interface.

Task complexityLong, multi-step objectivesShort, bounded tasks
Workflow ownershipAgent plans and carries the main taskUser keeps ownership; AI assists one moment
Interaction styleDelegation and progressInline suggestion and direct manipulation
Context awarenessCross-module project contextCurrent page, object, and user action
Expected outputA completed workflow or deliverableA result ready to review and apply

04 · Insight 02

AI should disappear into the workflow.

A Lite Agent should not feel like a smaller chatbot placed beside the product. Its value comes from appearing exactly where the user needs help, preserving context, and returning control immediately.

01

Inline, not adjacent

Place assistance inside the task instead of opening a separate conversation.

02

Do not interrupt

Keep the page, filters, selections, and scroll position intact.

03

Context is the entry point

Use the current creator, brief, project, or asset so users do not repeat setup.

04

Patterns should repeat

Make review, regenerate, apply, and escalation behaviors consistent across features.

05 · Insight 03

Trust comes from controllability.

The Project Review Lite Agent made this principle concrete. Trust did not come from presenting the model as more capable; it came from making every recommendation inspectable, reversible, and optional.

PROJECT REVIEW · CONTROL PATTERN

Make this project brief clearer?

The suggestion preserves the original intent while making the requirements more actionable for creators.

The user stays in control because AI proposes a change; it never silently becomes the source of truth.

06 · Behind the workshop

A short path from discussion to a shared interaction model.

01Workshop
02Cross-functional discussion
03Concept exploration
04Prototype
05Design review
06Final presentation

07 · Reflection

From one assistant to an ecology of AI interactions.

Old thinkingOne AI assistant should solve everything.
→
New thinkingDifferent AI forms should coexist.
Old thinkingConversation first.
→
New thinkingWorkflow first.
Old thinkingAI replaces interfaces.
→
New thinkingAI augments interfaces.

This workshop changed the way I think about AI-native products. The strongest experience is not always the most autonomous or conversational. It is the one that gives the right form of intelligence the right amount of responsibility at the right moment.

Related projectAuto Collaboration FeatureView Project →