Multi-Agent Architecture

The most advanced Claude integration pattern, deployed in apibstudyguide.

The 6-Agent Pipeline

Agent	Role	Model	Behavior
planner	TPM/PM — picks tickets, orchestrates	Opus	Never writes code. Analytics-driven scoring (Tier 1-4).
autoresearch	Competitive analysis	Sonnet	WebFetch/WebSearch. Creates tickets for gaps. Budget-limited.
creator	Subject expert	Sonnet	Creates/updates MDX content. Requests /simplify after changes.
validator	Static analysis	Sonnet	Runs content-validate.mjs and smoke-test.mjs.
reviewer	Factual accuracy	Sonnet	Read-only. Creates tickets for errors found.
ux-reviewer	Visual QA	Sonnet	Headless browser checks via ux-review.mjs.

Automated Pipeline (via SubagentStop hooks)

creator → /simplify → validator → reviewer → ux-reviewer

Each agent’s completion triggers the next. No human intervention required between steps.

Quality Gate (PreToolUse hook)

Tickets cannot be closed unless a UX review report exists that is:

• Less than 30 minutes old
• Shows 0 issues

This prevents regressions without requiring human vigilance.

Model Assignment Pattern

Opus for planning, Sonnet for execution — mirrors a PM/engineer split. The planner needs judgment and prioritization; execution agents need speed and accuracy.

Supporting Infrastructure

• tkt MCP server: Native ticket management within Claude sessions
• sync-feedback.mjs: SessionStart hook pulls student feedback from GitHub Issues
• Session logging: planner and autoresearch maintain .claude/*-log.json

Known Failure Mode: Agent Domain Blindness

When agents work outside their original context (e.g., English → French, IB → AP), they carry over assumptions invisibly. The pipeline’s layered review doesn’t catch these because no layer has the domain knowledge. Fix: research → configure → validate → regress-test. See agent-domain-blindness for the full pattern and source-agent-foreign-language for the case study.

See also: source-claude-usage-patterns, claude-integration-timeline, source-ib-competitive-landscape, agent-domain-blindness, source-agent-foreign-language