Project
TL Cloud Agents
Autonomous Claude agents for security scanning and error analysis
What It Is
Cloud agents that run 24/7 on Render, monitoring Traders Launch infrastructure. They use Claude Opus 4.5 with extended thinking as the orchestrator, with Gemini Flash for grunt work (context gathering, filtering, summarization).
The pattern: Opus decides, Gemini does legwork. This keeps costs down while maintaining quality for decisions that matter.
Architecture
Built on the Claude Agent SDK with a ClaudeAgent base class and selective tool mixins — CodeTools, LinearTools, DiscordTools, GeminiTools. Each agent gets an agent_context field that serves as its CLAUDE.md equivalent.
Gemini sub-tasks are handled via gemini_analyze tool — cheap analysis that doesn't need Opus-level reasoning.
What They Do
- 🔍Security scanning — Continuous audit of codebase for vulnerabilities
- 🚨Error analysis — Parse Sentry errors, identify root causes, create Linear tickets
- 📊Anomaly detection — Monitor PostHog metrics for unusual patterns
- 💬Discord updates — Post summaries and alerts to team channels
Why Build This
Manual monitoring doesn't scale. I can't watch dashboards all day, and neither can my team. But an agent can run continuously, notice patterns humans miss, and escalate only what matters.
The goal is getting to a place where most operational issues get noticed and triaged before anyone has to think about them.