Multi-Agent Parallel Build#

Spawn multiple coding agents in parallel to build independent components simultaneously. Reduces wall-clock time 3-5x for multi-page builds.

When to Use#

• Building 4+ UI pages that share a shell but have independent content
• Creating multiple API endpoint groups
• Any project with clearly separable, independent work units

Wave Architecture#

text

Wave 0 (Sequential — YOU do this):
  → Shared infrastructure: API, shell components, CSS, data layer

Wave 1 (Parallel — AGENTS do this):
  → Agent A: Page/Feature 1
  → Agent B: Page/Feature 2
  → Agent C: Page/Feature 3
  → Agent D: Page/Feature 4
  → Agent E: Page/Feature 5

Wave 2 (Sequential — YOU do this):
  → Integration fixes, cross-component wiring, testing

Step 1: Wave 0 — Build Shared Infrastructure#

Before spawning agents, create everything they'll all need:

1. API layer with all endpoints agents will consume
2. Shared shell (nav, header, theme, CSS variables)
3. Component library (cards, charts, tables, modals)
4. Data contracts (JSON shapes agents should expect from API)

This is critical — agents that build against a shared shell produce consistent UIs. Agents that each invent their own shell produce chaos.

Example shared shell:

javascript

// shell.js — agents import this
function createShell(pageTitle, navItems) { /* sidebar + topbar */ }
function createCard(title, content) { /* themed card component */ }
function mcFetch(endpoint) { /* API wrapper with base URL */ }

Step 2: Plan Agent Assignments#

Each agent gets:

• One clear deliverable (a page, a feature, a service)
• Path to shared resources (shell.js, shell.css, API base URL)
• API contract (which endpoints to call, expected response shapes)
• No overlapping files — each agent writes to its own directory/files

Write a task prompt for each agent:

text

Build {page_name} at {file_path}.
Import shared shell from {shell_path}.
Fetch data from these API endpoints: {endpoints}.
Expected data shapes: {json_examples}.
Use Chart.js/D3 for visualization. Dark theme. No frameworks — vanilla JS + HTML.

Step 3: Spawn Agents#

Use sessions_spawn or coding-agent skill to launch all agents simultaneously:

text

Agent A: "Build agents.html — display 215 AI agents in searchable grid..."
Agent B: "Build skills.html — display 197 skills with category filters..."
Agent C: "Build knowledge.html — display 6.8K knowledge records with search..."
Agent D: "Build tools.html — display 231 tools grouped by MCP server..."
Agent E: "Build workflows.html + archetypes.html — two pages..."

Key flags:

• Each agent gets its own working directory or clearly separate files
• Include the shared shell path and API contract in every prompt
• Set reasonable timeouts (10-20 min per page)

Step 4: Wave 2 — Integration Fixes#

Common issues after parallel build:

Double-prefix bug#

Agents often misconstruct API URLs: /api/mc/api/mc/ instead of /api/mc/. Fix with sed:

bash

sed -i '' "s|/api/mc/api/mc/|/api/mc/|g" static/mc/*.html

Inconsistent mcFetch usage#

Some agents inline fetch() instead of using the shared mcFetch(). Standardize.

Missing StaticFiles mount#

Server needs routes for the new static directories:

python

app.mount("/static/mc", StaticFiles(directory="static/mc"), name="mc-static")

Large payload timeouts#

If tables have massive text columns (system_prompt, etc.), agents may have built queries that select everything. Fix API to use lightweight selects.

Principles#

1. Wave 0 quality determines Wave 1 success — spend 60% of effort on shared infrastructure
2. No file overlap between agents — if two agents touch the same file, one will clobber the other
3. API-first — build and test all API endpoints before spawning UI agents
4. Expect 2-3 integration bugs per agent — budget 20 min for Wave 2 fixes
5. 5 agents is the sweet spot — more than 7 creates coordination overhead that exceeds time savings