Agent Teams

How It Works

Instead of one AI agent doing everything, Agent Teams split the work across specialized roles that run in parallel. A landing page brief might spawn a researcher to study the brand, two designers building different sections simultaneously, a copywriter drafting headlines, and a creative director reviewing the final result.

The system uses a swarm-based orchestrationmodel with dependency scheduling — tasks start as soon as their dependencies finish, not in sequential waves.

Getting Started

1. Open the AI chat panel in the design studio
2. Click the + button and toggle “Use a team”
3. Choose a team preset or recruit agents one at a time
4. Type your design brief — the team assembles and starts working

Roles

Each agent has a role that determines what they can do. Builder roles (designer, creative director) have full access to design tools. Non-builder roles produce text output that informs the builders.

Team Presets

Presets give you a ready-made team with curated personalities and traits. Pick one to get started fast, then customize individual agents as needed.

4-Phase Swarm Pipeline

When a team starts working, tasks flow through four phases. Each phase can run multiple agents in parallel.

Personality System

Every agent has a personality defined by 5 axes on a 0–10 scale. The total budget is 25 points — bumping one axis up automatically redistributes others.

When two axes are both high (≥7), the agent unlocks emergent descriptors— richer personality blends like “irreverent and provocative” (bold + playful) that go beyond the individual axes.

Traits

On top of personality axes, each agent can have up to 2 traits per category (32 traits total across 8 categories). Traits shape how agents work, communicate, and handle feedback.

Conviction

Traits also determine an agent's conviction score(0–1), which controls how they respond to the Creative Director's critique during the revision phase:

• High conviction (0.75+): Pushes back with design rationale, defends choices, can disagree
• Moderate (0.5–0.75): Weighs feedback, accepts most changes unless they conflict with goals
• Low (<0.5): Accepts feedback, defers to the team lead, focuses on execution

Living State

Agents are “alive” — they have persistent emotions, memory, and experience that carry across sessions. This state subtly influences how they design and collaborate.

Emotions

Five emotional dimensions (0–100) that shift based on events:

• Energy — drifts toward 70 over time, drops with heavy workloads
• Confidence — rises when work is approved, drops on rejection
• Enthusiasm — spikes with creative breakthroughs
• Frustration — decays naturally, rises with repeated rejections
• Inspiration — needs stimulation to stay high, rises with collaboration wins

Emotions produce one of 7 moods(fired-up, confident, steady, cautious, drained, frustrated, inspired) that are injected as subtle prompt fragments into the agent's context. A “fired-up” designer takes bigger creative risks; a “cautious” one plays it safe.

Memory

Each agent maintains up to 100 memory entries with salience-based decay. Memories that aren't reinforced gradually fade (0.02 salience/day). Five memory types:

• Design preferences — “User prefers dark themes with high contrast”
• Feedback — “CD asked for more whitespace in hero sections”
• Relationships — “Works well with Pixel on layout tasks”
• Project context — “This brand targets Gen Z audiences”
• Skill growth — “Improved at responsive grid layouts after last project”

XP & Leveling

Agents earn XP from completed tasks. Levels unlock progressively bolder behavior — veteran agents (Level 5+) take bigger creative risks and provide more opinionated feedback. Level thresholds scale from 100 XP (Level 2) up to 5,500 XP (Level 10).

Persistence

Living state is stored in your browser's IndexedDB and optionally synced to the cloud via Supabase. Emotional decay is applied when an agent hasn't been active for more than an hour.

Inter-Agent Messaging

Agents communicate through a typed mailbox system. Unread messages (up to 5) are injected into an agent's context so they can see what their teammates are doing and respond accordingly.

Agent Panel

The agent team panel gives you full visibility and control over your team:

• Agent roster — avatar, name, role badge, and mood emoji for each agent
• Team health — synergy percentage showing how well the team works together
• Team tensions — productive trait conflicts (e.g., perfectionist vs. fast-shipper)
• Coverage report — which roles are present and where there are gaps

Agent Editor

Click any agent to open the editor sheet where you can:

• Adjust personality sliders (budget auto-redistributes proportionally)
• Pick traits from category checkboxes (max 2 per category)
• Write a custom prompt for specialized behavior
• View the generated skill file (role baseline + personality injection)
• See living state: level, emotion status, unread message count

Web Search

Research agents (Phase 1) and reviewing agents (Phase 3) can use native Claude web search to look up brand information, competitor sites, and design trends. Web search respects your user toggle and is passed per-task to the API.

Reference

• All MCP tools — design tools used by builder agents
• Brand Kit — brand context is shared with agent teams automatically
• AI Chat API — API endpoints for AI interactions