Multi-Agent Design Patterns for Belonging — Case Study

The problem space

Belonging is human labor. What happens when AI enters the room?

Online communities are where millions of people weather crises, form real relationships, and build a sense of belonging with people they may never meet. That belonging isn't automatic — it's produced by labor: members welcome newcomers, check in on each other, amplify each other's creative work, and protect the space from harm.

Generative AI is now entering these spaces. The dominant instinct is to automate — auto-answer the newcomer, auto-generate the art, auto-moderate the conflict. But every one of those acts is also a moment where belonging is made. Automate it away and you may deliver the information faster while quietly removing the human contact that produced connection in the first place.

The design challenge: how can advanced AI be designed to support — rather than displace — the human practices through which belonging is produced? Not "should AI be here," but "what role should it play, and where would that role do more good than harm?"

Scoping the problem

From an unbounded question to three answerable studies

"How should AI support belonging" is too big to design against directly. I scoped it into three sequential studies, each answering a question the next one needed — moving from human dynamics, to AI-system reality, to a synthesis that bridges the two.

Study 1 · Understand the human mechanismHow does supportive communication actually create belonging in a thriving online community? (Diary study + interviews with the BTS/ARMY fandom.)

Study 2 · Understand the design materialHow do practitioners who build multi-agent generative AI systems conceptualize, structure, and constrain them? (Interviews with AI practitioners at Microsoft.)

Study 3 · Synthesize into design movesWhat AI agent roles do those human mechanisms imply — and under what conditions do they help rather than harm? (Research-through-design synthesis → the pattern library.)

Why this scoping matters as design: I deliberately refused to start at the interface. The patterns are grounded in why belonging forms before deciding what an agent should do — so each design move traces back to evidence, not assumption.

Research-through-design

How three studies became a design library

Research-through-design treats the act of designing as a way of producing knowledge. Each study generated insight; the synthesis turned that insight into conceptual design moves — named agent roles with behaviors, instantiations, risks, and diagnostic questions a real team could apply.

STUDY 1

Get close to the human truth

Diary study + semi-structured interviews with BTS/ARMY members surfaced the real mechanisms of belonging: onboarding vulnerability, the social cost of disclosing struggle, under-recognized creative labor, and threats to community integrity.

Diary studyInterviewsThematic analysis

STUDY 2

Understand the AI material

Interviews with AI practitioners building multi-agent systems revealed how they think (org metaphors, "team of experts," "divide & conquer"), two collaboration paradigms (AI-dominant vs AI-assisted), and cross-cutting risks: hallucination, error propagation, the black-box problem.

Practitioner interviewsMulti-agent systemsTrust & transparency

STUDY 3

Synthesize into patterns

For each human insight I asked: what agent role does this imply? Role concepts were clustered by the community challenge they addressed, producing nine patterns in four themes — four specified in full depth, each with behaviors, risks, and fit questions.

RtD synthesisPattern languageVignettes

human mechanism → AI design material → transferable, evidence-grounded patterns

Analytical rigor

How the practitioner data became evidence

The AI-practitioner study (Study 2) wasn’t summarized from memory — it was analyzed systematically. For each participant, I built cognitive maps organized by interview-question type, then extracted the concepts embedded in their answers. Across all participants, I grouped semantically similar concepts — question type by question type — into themes, so every theme stays traceable back to the specific participant data it came from.

Data analysis process: per-participant cognitive maps organized by interview question type with extracted concepts (left), then thematic analysis grouping similar concepts across all participants into themes (right). — **Left:** per-participant cognitive maps, organized by interview-question type, with concepts extracted from each participant’s data. **Right:** thematic analysis — semantically similar concepts grouped across all participants into themes for each question type.

Why it matters: this two-step pipeline — individual cognitive mapping → cross-participant theming — is what let the synthesis (Study 3) ground each design pattern in evidence rather than impression.

The outcome of the research

Four primary agent patterns

Each pattern names a role an AI agent might play, the belonging challenge it addresses, the coordinated sub-agents that carry it out, and — critically — the risks and diagnostic questions that decide whether it belongs in a given community. A guiding constraint runs through all four: amplify and protect human connection; never substitute for it.

Theme 1 · Lowering barriers to participation

Newcomer Bridge Agent

Newcomers need to learn norms and find the right people, but asking basic questions exposes their outsider status. The agent privately learns their goals and bridges them to the people who can welcome them — it never befriends them itself.

Sub-agents

User ProxyKnowledge CuratorGuide WriterContribution Coach

Key risk: pre-empting the welcome question can remove the moment of human contact that produces belonging.

Theme 2 · Emotional & social support

Support Companion Agent

Members carry real emotional weight but hesitate to disclose publicly. The agent offers low-pressure, dignity-preserving check-ins with a consent-based escalation model — never therapy, never imposed, always exitable.

Core moves

Gentle check-in"No response needed"Resource routingConsent escalation

Key risk: subtle care can feel like surveillance; false positives erode trust fast.

Theme 3 · Expression & collective sense-making

Creative Signal Booster Agent

Creative work builds belonging but gets drowned out. The agent helps members share and be seen — captions, translation, accessibility, amplification — while boosting signal, not generating content (calibrated to each community's authorship norms).

Core moves

Caption / translateAlt-text by defaultTargeted amplificationSupportive feedback

Key risk: can homogenize style and reinforce popularity hierarchies.

Theme 4 · Protecting community integrity

Boundary & Integrity Agent

Belonging needs protection from harm. The agent does quiet boundary work — detecting scams and coordinated harassment, then handing evidence and recommendations to human moderators with a built-in bias check. It supports moderators; it never enforces autonomously.

Sub-agents

DetectorThreat IntelligenceBias CheckExplainer / Writer

Key risk: false positives disproportionately hit non-native speakers; surveillance chills participation.

Mapping all nine patterns to the member journey

These four are only part of the picture. This interactive journey map shows how all nine agent patterns — including cross-cutting ones like the Boundary & Integrity Agent and the Identity Mask & Credit Keeper — align to each stage of a member's journey, from newcomer to experienced member. Click through the stages to see which agents are active when.

Member Journey — AI Agent Design Patterns for Community Belonging Open full screen ↗

The patterns, vibe-coded as designs

See each agent in the room

To pressure-test the patterns as design, not just theory, I vibe-coded each one into a working, deployed app — a realistic community interface that walks through the exact vignette from the research. Open it below, pick an agent, and step through the scenario to watch how the agent works at the margins — private, consent-first, human-in-the-loop.

Designed with AI (vibe-coded). I built this as a live, deployed app by describing the vignettes to AI coding tools and iterating in natural language — turning a 160-page dissertation into a working product prototype. The workflow itself is the artifact: it shows designing with AI, not just about it.

Multi-Agent Design Patterns — live, vibe-coded app Open the live app ↗

Live app, deployed on Google Cloud Run. Scenarios are fictionalized composites — the underlying dynamics are drawn directly from the dissertation's research data and vignettes. If the embed doesn't load, use “Open the live app ↗”.

How design principles are woven in

Every foundational principle from the research is woven directly into the interactive mechanics and visual cards of the live design. Here is exactly how and where each one shows up in the interface:

01Agents act privately & consent-first

Support Companion — Mina's silence check-in: when Mina goes quiet for ~10 days, the agent reaches out privately in DMs rather than calling her out publicly, offering a low-pressure "🌿 I'm okay, thanks / Just browsing" option — she is never forced into a social interaction or penalized for resting.
Newcomer Bridge — Jane: Jane can choose to cross the warm bridge or decline quietly. The agent initiates support one-to-one, so the user keeps complete control of their visibility.

02Agents amplify & route — never replace humans

Creative Signal Booster — Mina's art booster: the agent strictly avoids "AI-slop" content generation — it does not write about or critique her artwork. Instead it gives structural leverage: auto-generated accessibility ALT-text so visually-impaired fans can enjoy her style, and routing the post to the dedicated #lyric-analysis-deep-dive community so the signal reaches interested human minds. Mina remains the sole owner, creator, and voice.

03Human-in-the-loop moderation — evidence brief & bias check

Boundary & Integrity — moderation co-pilot: the agent is strictly prohibited from instant, autonomous bans. It serves as a co-pilot to human moderator @RM_Lead_Mod, compiling a structured evidence brief that matches timestamps, patterns, and severity.
Style-bias safety warning: a dedicated card flags that bilingual member @JiminLove_KR — though typing rapidly about ticket trades — is an authentic, non-native-speaking human. The agent explicitly prompts the moderator to exclude them from quarantine, preventing algorithmic bias against cultural and linguistic variation.

04Continuous, clear transparency notes

Every time an agent produces a layout or assistance card in a DM or channel, it carries an explicit, bottom-anchored 🔍 Transparency Note clarifying exactly what data was analyzed (e.g., participation-frequency metadata only, join timestamps, image-contrast tags) and the architectural boundaries protecting user privacy.

Outcome & contribution

A blueprint teams can actually use

The nine patterns consolidate into a transferable design blueprint and an operational meta-playbook for community designers, platform architects, and researchers — pairing empirically grounded theory with diagnostic questions that help a team decide which pattern fits, and where automation would do harm. The dissertation makes two theoretical moves and delivers two practical tools.

Two theoretical moves, two practical deliverables

Belonging-as-Labor theory

Re-frames belonging as a dynamic practice that demands creative, infrastructural, and emotional labor from members — so AI should amplify that labor, never automate it.

Social-Metaphor Tension theory

Names the gap between the human metaphors practitioners design AI with (team, assistant) and the social accountability — empathy, negotiation — that current AI can't actually fulfil.

Design Blueprint deliverable

A transferable set of 9 multi-agent patterns across 4 themes, bridging real community needs with technical multi-agent architectures.

Operational Meta-Playbook deliverable

A six-step process for taking patterns from paper to production — diagnosing belonging breakdowns before prescribing any AI intervention.

The meta-playbook — 6 steps from paper to production

01Diagnose belonging

Diagnostic questions across five belonging dimensions map to the thematic groups — no prior taxonomy knowledge needed.

02Select patterns

Compose patterns to fit (e.g. Newcomer Bridge + Boundary Agent); they're designed to be combined.

03Instantiation strategy

Re-instantiate per community culture, scale, and risk tolerance — adapt, don't copy.

04Agent configuration

Decide single-agent vs. multi-agent architecture for the specific community need.

05Transparency & consent

Pre-deployment protocol: opt-out options, pause features, visible human-escalation paths.

06Evaluate care risks

Assess failure modes — over-automation, loss of implicit care, power imbalances.

Reflection

What this case demonstrates

Framing an ambiguous problem space, scoping it into answerable research, synthesizing evidence into a reusable design system of patterns, and translating abstract concepts into concrete, interactive UI — end to end, by one designer.

What I'd do next

Move from blueprint to evidence: deploy coordinated patterns in live communities, instrument belonging and trust signals, extend the pattern language beyond fandom (gaming, health, education), and study the longitudinal effects of AI-mediated belonging.

Published research

DIS 2025

Designing with Multi-Agent Generative AI: Insights from Industry Early Adopters

CHI EA 2025

Early Adopters' Use of AI-Driven Multi-Agent Systems

C&C 2024

From Digital Practices to Bond Formation: BTS Online Fandom