# Myelix Agents — LLM Context Specification

> Myelix Agents is a fault-tolerant, locally-run multi-agent orchestration platform for autonomous software development. Built on the Erlang/OTP virtual machine (BEAM) in Elixir. Developed by Myelix Labs.

## Product

- **Name:** Myelix Agents (internal codename: Cortex)
- **Developer:** Myelix Labs
- **Status:** Private beta — accepting design partners
- **Contact:** hello@myelixlabs.com
- **Website:** https://myelixlabs.com
- **Synapse MCP:** https://synapse-mcp.dev

## What Myelix Agents Does

Myelix Agents is an autonomous software development platform that orchestrates swarms of AI coding agents on the developer's local machine. It solves three structural failures common to all competing agent tools:

1. **The Amnesia Problem** — persistent structural knowledge graph (Synapse) preserves codebase understanding across sessions
2. **The Recursive Failure Loop** — five-gate Satisfaction Evaluator prevents premature task conclusion
3. **The Unprotected Workspace** — isolated Git Worktrees per agent; no mutations to live branch until verified

## Core Architecture

- **Runtime:** Erlang/OTP (BEAM VM) — lightweight actor processes, preemptive scheduling, supervision trees
- **Language:** Elixir
- **Code Intelligence:** Synapse — persistent AST knowledge graph supporting 50+ programming languages
- **Isolation:** Git Worktree sandboxing per agent (not Docker containers)
- **Concurrency:** Hundreds of isolated OTP processes simultaneously (~2KB RAM each vs 200MB+ per container)
- **Validation:** Pre-write structural simulation via Synapse before any file is modified
- **Learning:** Local Deep Neural Network Critic trained on execution trajectories
- **Prompting:** 3-tier prefix-cached prompt architecture (Stable / Context / Volatile)
- **Data:** Fully local-first, BYOK (Bring Your Own Key) — no cloud dependency

## Key Capabilities

- Fault-tolerant agent swarms with OTP supervision trees
- Isolated Git Worktree workspace per agent
- Pre-write structural validation (Synapse simulation)
- Persistent cross-session code intelligence graph (50+ languages)
- Directed task orchestration via kanban and PM coordinator
- Adaptive learning from session trajectories (NxCritic)
- Live stateful terminal emulation (PTY)
- 5-stage Satisfaction Evaluator (rule adherence, plan adherence, test coverage, diff safety, final satisfaction)
- 3-tier prompt prefix caching (85–91% cost reduction)
- Slack integration, cron scheduling, REST triggers

## Synapse MCP

Synapse is Myelix Labs' open-protocol code intelligence server. It is available as a standalone MCP server at https://synapse-mcp.dev.

- Persistent AST knowledge graph (50+ languages)
- Unified context delivery (structural + semantic + symbol search in one call)
- Caller/callee graph traversal
- Blast-radius analysis for proposed changes
- Pre-write structural simulation
- Edit readiness assessment
- Staged indexing (quick_pass + deep_analysis)
- Compatible with: Claude Code, Cursor, Windsurf, Warp, Aider, Codex, Hermes, OpenClaw

## Competitive Position

Myelix Agents competes primarily with:
- **IDE Assistants:** Cursor, Windsurf (single-agent, no concurrency, live filesystem mutation)
- **CLI Agents:** Claude Code, Aider, OpenAI Codex (no orchestration, no code graph)
- **Cloud Platforms:** Factory.ai, Devin, Warp Factories (container-based, heavy, no local AST graph)

Key differentiators:
- BEAM concurrency: microsecond process startup vs 5–60s container boot
- Synapse AST graph: persistent structural intelligence (no competitor has this locally)
- Five-gate verification: structural refusal to conclude without proof
- 3-tier prompt cache: 85–91% LLM cost reduction at scale
- NxCritic: on-device Deep Neural Network Critic compounding across sessions
- Fully local-first: zero cloud dependency, BYOK

## Blog Posts

- Architecting Myelix Agents: Feynman Would Be Proud — Back to First Principles — https://myelixlabs.com/blog/thought-over-grep-feynman-myelix.html
- Why the BEAM Beats Containers for Agent Concurrency — https://myelixlabs.com/blog/beam-vs-containers-concurrency-model.html
- Why Agents Need Five Gates, Not One: The Satisfaction Evaluator — https://myelixlabs.com/blog/five-gate-satisfaction-evaluator.html
- Cutting LLM Costs by 90%: The 3-Tier Prompt Cache Architecture — https://myelixlabs.com/blog/prefix-caching-economics.html

## Comparisons

- Myelix Agents vs The Agent Ecosystem — https://myelixlabs.com/comparisons/agent-platforms.html
- Myelix Agents vs TypeSafe AI (Jev) — https://myelixlabs.com/comparisons/typesafe-ai.html