Architecture

Three concurrent polling loops run continuously:

• Message Loop (2s): Handles incoming messages (WhatsApp monitoring/triggering)
• IPC Watcher (1s): Picks up result files from IPC_DIR/{id}/results/
• Scheduler (60s): Checks SQLite for due tasks and fires agent invocations

Container isolation via Apple Container (macOS Tahoe+) or Docker. SQLite state persistence with two-cursor at-least-once delivery. Per-group memory via CLAUDE.md files.

Self-Continuation Pattern

The key pattern that enables autonomous operation: agents call schedule_task() to queue their own future invocations. The scheduler loop fires the task when due. This means no cron jobs, no infinite loops, no external orchestration. Each agent: reads CLAUDE.md (state) then does web research, writes findings to IPC_DIR, and schedules its next run. Clean, resumable, fault-tolerant.

Supabase Push Pattern

IPC watcher picks up JSON result files from agents. A post-processor normalizes the output and executes a Supabase INSERT into geopolitical_findings (or extends content_items). The webapp already reads from Supabase for all research data. Adding another source table is trivial.

Llama 4 Scout

Cost: Free (self-hosted). Quality: Decent for classification and filtering tasks. Can handle binary relevance scoring (is this article about geopolitics? yes/no). NOT suitable for deep geopolitical analysis requiring nuanced reasoning about second-order effects, probability weighting, or market transmission mechanisms.

MiniMax

Cost: API pricing competitive with GPT-4.1 Mini. Quality: Chinese model, strong at language tasks and translation. Significant quality gap vs Opus/GPT-5 for multi-step reasoning. Potential geopolitical bias concerns when analyzing events involving China.