DEV Community

Domonique Luchin
Domonique Luchin

Posted on

CLAWKILLER: My 8-Agent LangGraph Autonomous Ops Platform

CLAWKILLER is the name I gave my AI orchestration system.

It runs six businesses from a single server. Here is the architecture.

Why I Built It

I could not afford to context-switch between six companies during a full-time engineering job. I needed a system that could intake a goal, plan execution, use tools, and complete the task without me sitting there.

LangGraph gave me the graph-based agent orchestration I needed. Supabase gave me persistent memory. Claude and GPT-4o gave me the reasoning layer.

The 8 Agents

SUPERVISOR — Command parser and router. Every task enters here first. Routes to the appropriate specialist agent based on intent.

BROWSER — Playwright-powered web agent. Account creation, form fills, web scraping, social posting.

CONTENT — Writes articles, captions, scripts. Pulls from session logs and build notes.

COMMS — Drafts emails, SMS, and Slack messages. Handles follow-ups and outreach sequences.

DATA — Queries Supabase, runs reports, pulls lead pipeline status across all six companies.

SCRAPER — Nightly lead generation. Hits county deed records, filters by acquisition criteria, logs to CRM tables.

SCHEDULER — Creates calendar events, queues time-sensitive tasks, manages approval gates.

AUDIT — Reviews completed tasks for errors, logs anomalies, surfaces issues requiring human review.

Memory Architecture

Every agent writes state to Supabase after each step. Session continuity does not depend on keeping a single process alive.

The jarvis_memory table stores intent, execution history, tool results, and status. Any agent can query it to pick up where another left off.

Control Plane vs Intelligence Layer

CLAWKILLER is the control plane. Claude and GPT-4o are the intelligence layer.

CLAWKILLER decides what runs and when. The LLMs decide how.

Current Status

Schema live on Supabase (project: whxtjboruayowkqyjvcq). Services running on lb-telecom-01 ports 8000, 3001, and 3000. Claude Code installed for dispatch via the claude_code_queue table.

This is not a demo project. It runs production workloads every night.

Top comments (0)