THE FIRM OS

Back to Dashboard

What if your startup had aSelf-Organizing AI Executive Team?

A purpose-built multi-agent OS. Zero frameworks. One Edge Function. Four tables. They pull live production data, debate strategy, and run operations proactively so you can focus on leading.

4 Autonomous Agents
1 Edge Function
4 Database Tables
~300 Lines of Code
0 Frameworks

Architecture at a Glance

Telegram
Webhooks
Dashboard
Realtime
POST /nexus/{agent_id}
Single Orchestrator Edge Function (~300 LOC)
Load Context
→
LLM Execution
→
Handle Tools
→
Store & Deliver
Nexus DB
agents
messages
memory
activity_log
Production DB
Read-Only Data Bridge
profiles
subscriptions
conversations
x_interactions

Technical Deep Dive

Prompt-Driven Architecture

Agent quality lives in SOUL prompts, not code. Behavior changes are markdown edits, not redeployments. The orchestrator is completely generic; personality, action tiers, and coordination rules exist purely in prompt space.

Structured Multi-Agent Debate

Engineered disagreement, not consensus-seeking. The COO explicitly gathers conflicting perspectives via the contact_agent tool from domain agents before synthesizing a final recommendation for the CEO.

Context Engineering & Memory

No pgvector overhead. A tiered key-value memory powered by to_tsvector. Critical context (goals, decisions) is pinned to the edges of the prompt, while query-relevant memories are dynamically retrieved and budgeted to prevent the "Lost in the Middle" trap.

Live Production Data Bridge

A read-only dual-Supabase-client connects agents to the live business database. Via query_eddie, agents pull from 5 aggregated report types (e.g., MRR, RAG performance) to ground decisions in reality. No fragile webhooks.

Proactive Autonomous Ops

Powered by GitHub Actions. Agents wake up every 4 hours (Heartbeats) to check goals and flag anomalies. At 6 PM, the COO sends a Daily Summary. Friday at 5 PM, a Weekly CEO Brief. They surface what matters without being asked.

Zero-Framework Simplicity

No LangChain, no LangGraph, no message queues. One Supabase Edge Function handles Telegram, Dashboard, cron, and agent-to-agent communication. 4 tables replace the typical 11+. 64% less surface area for bugs.

Resilient Tool Execution

Pre-execution argument validation prevents the #1 failure mode in agentic systems (hallucinated parameters). If a tool fails, the Evaluator-Optimizer uses Structured Error Reflection to diagnose the root cause before attempting recovery.

Self-Improvement Flywheel

Using the SCOPE dual-stream framework, the system automatically analyzes failed interactions weekly. It extracts tactical session rules and proposes high-confidence strategic rules for SOUL prompts, guarded by an automated regression eval suite.

The Improvement Cycle

1
BASELINE
Agent queries live metrics
eddie:baseline:*
2
RECOMMEND
Suggests config change
Admin applies tweak
3
TRACK
Monitor over 48h
eddie:experiment:*
4
MEASURE
Compare vs. baseline
query_eddie
5
LEARN
Save pattern to memory
eddie:learning:*

The Agent Toolkit

contact_agentCoordination

Synchronous peer communication with Call Depth tracking

write_memoryMemory

Persistent knowledge storage (goals, decisions, learnings)

read_memoryMemory

Postgres tsvector full-text search over keys and values

delete_memoryMemory

Prune obsolete state (e.g., resolved experiments)

query_eddieExternal

Live production data bridge (5 structured report types)

web_searchExternal

Tavily API search for competitor intel and research

fetch_github_commitsExternal

Grounds engineering updates in actual codebase diffs

fetch_stripe_mrrExternal

Raw MRR retrieval (often supplemented by query_eddie)

update_statusState

Global state updates visible to CEO Dashboard and peers

read_statusState

Pull peer state (avoids burning tokens asking 'what are you doing?')

escalateAlert

Throws CEO alert to Dashboard and Telegram immediately

set_heartbeat_intervalState

Dynamic adjustment of cron wake-ups based on crisis/quiet

Meet The Partners

Elon Eddie

Chief Operating Officer ⚡

The right hand. Synthesizes cross-functional decisions, maintains quarterly goals, and coordinates the entire partnership.

Key Tools
contact_agentwrite_memory (decisions)query_eddieescalate
Proactive Duties
  • Structured Debate Synthesis
  • Daily Summaries (6pm)
  • Weekly CEO Brief (Fri 5pm)

Egor

VP Engineering ⚙️

The architect. Handles technical architecture, tech debt, deployment timelines, and engineering resources.

Key Tools
fetch_github_commitsquery_eddie (system_health, alerts)read_memory
Proactive Duties
  • Tech Debt Tracking
  • System Degradation Alerts
  • Architecture Reviews

Finn

VP Finance 💰

The bottom line. Tracks runway, MRR, cash flow, and payment tracking. Always keeps the budget in focus.

Key Tools
fetch_stripe_mrrquery_eddie (business_overview)contact_agent
Proactive Duties
  • Runway Modeling
  • MRR Baseline Tracking
  • Resource Allocation Approval

Oz

VP Operations 🏗️

The logistics master. Manages vendor relations, compliance calendars, customer feedback patterns, and day-to-day operations.

Key Tools
query_eddie (x_engagement, conversation_quality)write_memory (learning)
Proactive Duties
  • X Growth Loop
  • Customer Health Analysis
  • Compliance Calendar Management
1
Edge Function
(vs 6 in v1)
4
Database Tables
(vs 11 in v1)
300
Lines of Orchestrator
Code
-64%
Surface Area
For Bugs