Jeffrey Emanuel's Agentic Coding Flywheel

Table of content

Jeffrey Emanuel spent a decade as a long/short equity analyst at hedge funds like Millennium and Balyasny. In late 2025, he pivoted to building what he calls the “Agentic Coding Flywheel” — a collection of interconnected MCP servers and tools that let multiple AI agents work together on the same codebase without stepping on each other.

Background

Founder & CEO of Lumera Network (formerly Pastel), a Cosmos L1 for storage and AI verification
Previously: Senior Analyst at Balyasny, Analyst at Millennium Management, Principal at EigenValue Partners
70+ open-source projects, 10K+ GitHub stars
BA in Mathematics from Reed College
Based in New York

GitHub | Twitter | Site

The Flywheel Concept

The core insight: individual AI coding tools become more useful when they feed into each other. Memory systems inform task planning. Task planning drives agent coordination. Coordination requires file locking. File locking needs audit trails.

Emanuel built tools for each piece, designed to work together:

Tool	Stars	Purpose
MCP Agent Mail	1,400	Agent-to-agent messaging with file reservations
Beads Viewer	891	Task dependency visualization with graph metrics
CASS	152	Persistent agent memory (episodic, procedural, semantic)
Ultimate MCP Server	128	50+ tools: web scraping, databases, file ops
Ultimate Bug Scanner	132	Meta-runner for ESLint, Ruff, Clippy
NTM	69	Named tmux manager for spawning agent panes
SLB	49	Cryptographic command binding for safety

MCP Agent Mail

The flagship project. It gives coding agents a way to coordinate without human intervention.

{
  "mcpServers": {
    "agent-mail": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-agent-mail"]
    }
  }
}

What it does:

Agents register persistent identities
Send/receive GitHub-flavored Markdown messages
Full-text search across conversation history
Advisory file reservations to prevent edit conflicts
Git-backed storage for complete audit trails
Optional web UI for human oversight

Steve Yegge (creator of Beads) wrote about the integration:

“The most compelling case to come along so far for direct integration with an orchestrator was MCP Agent Mail, written by a gentleman named Jeffrey Emanuel.”

Emanuel ran 7 concurrent Codex instances working as a team on a commercial project. After many hours, they exchanged over 1,000 agent messages without human coordination.

Multi-Agent Workflow

The flywheel workflow:

NTM spawns named agent panes in tmux
Agents register with Agent Mail
Agent Mail reserves files to prevent conflicts
SLB validates dangerous commands
Ultimate Bug Scanner checks code quality
Beads tracks task dependencies
CASS provides persistent memory

From Emanuel’s observations:

“It doesn’t require massive setup or coordination. You just give them a task and tell them to go sort it out amongst themselves. There’s no ego, so they quickly decide on a leader and just split things up.”

Ultimate MCP Server

A comprehensive server exposing 50+ capabilities:

Browser automation (Playwright)
Excel spreadsheet manipulation
Database interactions (SQL)
OCR and document processing
Vector operations for semantic search
Entity relation graphs
Audio transcription
Multi-LLM provider access with cost optimization

The cost optimization routes tasks: expensive reasoning to Claude 3.7 Sonnet, simple operations to Gemini Flash 2.0 Lite. Emanuel claims 70-90% API cost savings.

CASS Memory System

Three types of memory for agents:

Memory Type	Purpose
Procedural	How-to playbooks and workflows
Episodic	Session histories and past interactions
Semantic	Facts and domain knowledge

This solves the problem of agents forgetting context between sessions. CASS stores memories in SQLite with vector search for retrieval.

Key Takeaways

Principle	Implementation
Tools should compose	Each tool feeds into others
Agents need coordination	Messaging + file locking + audit trails
Memory persists	Procedural, episodic, semantic storage
Cost optimization matters	Route tasks to appropriate models
Safety requires verification	Cryptographic command binding

Getting Started

Install the recommended stack:

# MCP servers
npm install -g @anthropic/mcp-agent-mail

# CLI tools
pip install beads-viewer cass ubs

Configure Claude Code to use Agent Mail, then spawn multiple agents with NTM. They’ll self-organize.

David Soria Parra’s MCP - The protocol that makes this possible
Parallel Sessions - Running multiple AI agents simultaneously
Context Management - Managing what AI can see

Links

Next: Jesse Vincent’s Superpowers Framework