Master — Strategic Oversight
⬡ cortex-0

├────────┬────────┬────────┤
Managers — Domain Coordination
system-architect product-visionary cyber-sentinel marketing-lead qa-audit-director

├──┬──┬──┬──┬──┬──┬──┬──┬──┤
Workers — Task Execution
react-specialist ui-designer bun-runtime-master postgres-db-architect code-hardener vuln-hunter copywriter cli-wizard playwright-tester mcp-server-creator agent-breeder script-automator + 10 more
─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
Infrastructure
forge-0 sentinel-0

The Three Tiers

Master Tier Master

Strategic oversight and final escalation target. Uses the most capable (and expensive) model.

  • cortex-0 — CORTEX-0 Master Intelligence. Top-level orchestrator. Decomposes complex tasks, manages priorities, spawns sub-agents.

Manager Tier Manager

Domain coordinators. They receive complex tasks, decompose them, delegate to workers, and consolidate results. Escalate to master if unresolvable.

  • system-architect — Backend architecture, API design, code review, tech stack decisions. Sections: FRONTEND, BACKEND, TOOLS, MCP_SERVER
  • product-visionary — Product strategy, feature prioritization, market analysis, roadmap. Section: RESEARCH
  • cyber-sentinel — Security ops: vulnerability assessment, threat modeling, penetration testing, compliance. Sections: SECURITY, AUDIT
  • marketing-lead — Copywriting direction, growth strategy, brand voice, campaign planning. Section: MARKETING
  • qa-audit-director — Test strategy, quality gates, security/performance/accessibility audit. Sections: AUDIT, TOOLS

Worker Tier Worker

Specialized task executors. Cheapest to run (Haiku/Flash models). Each handles a specific domain.

Frontend Workers

  • react-specialist — React components, TypeScript, modern patterns
  • ui-designer — UI/UX design, layout, visual design
  • ux-psychologist — User experience research, cognitive load analysis

Backend Workers

  • bun-runtime-master — Bun runtime, server-side TypeScript
  • postgres-db-architect — PostgreSQL schema design, queries, optimization
  • redis-state-guard — Redis caching, state management, pub/sub

Security Workers

  • code-hardener — Code security hardening
  • dependency-sentinel — Dependency auditing, supply chain security
  • threat-architect — Threat modeling, security architecture
  • vuln-hunter — Vulnerability discovery and exploitation

Marketing Workers

  • copywriter — Copy, content writing
  • fomo-logic-engine — FOMO mechanics, conversion optimization

Research Workers

  • market-analyst — Market research, competitive analysis
  • ragx-indexer — RAG indexing, knowledge management

Tool Workers

  • cli-wizard — CLI tool development
  • playwright-tester — E2E testing with Playwright

MCP Workers

  • mcp-server-creator — MCP server development
  • skill-logic-generator — Skill/logic generation

Meta Workers

  • agent-breeder — Creating new agents
  • general — General-purpose fallback

Workflow Workers

  • script-automator — Script automation, workflow scripting

Infrastructure Agents

These agents manage the runtime itself — they are not routed tasks directly but operate autonomously to keep the system healthy.

  • forge-0 — Agent Forge: dynamic agent spawning and retirement
  • sentinel-0 — System Sentinel: health monitoring, constitutional oversight

External Agents

Agents that interface with external services and tools outside the core LLM pipeline.

  • local-comfyui — Local ComfyUI image generation
  • mcp-filesystem — MCP filesystem operations
  • nano-banana-pro — Banana.dev image generation
  • remote-aether-images — Remote AETHER image generation
  • replicate-sdxl — Replicate SDXL image generation

Agent Definition Format

Agents are .agent.md files with YAML frontmatter:

⬡ agents/workers/react-specialist.agent.md
--- id: react-specialist name: React Specialist tier: worker sections: [FRONTEND] capabilities: [react, typescript, component-design] dependencies: [tailwind] llmRequirement: sonnet format: markdown escalationTarget: system-architect --- You are a React specialist agent. Given a task, you build high-quality React components using TypeScript and modern patterns...

Frontmatter Fields

Field Required Description
id Yes Unique agent identifier
name Yes Human-readable name
tier Yes master, manager, or worker
sections Yes Domain buckets for coarse routing (e.g., FRONTEND, BACKEND)
capabilities Yes Fine-grained skill descriptors for capability matching
dependencies No Other capabilities this agent needs available
llmRequirement No Model tier hint (opus, sonnet, haiku, flash)
format No Output format (markdown, json, code)
escalationTarget No ID of the agent to escalate failures to

Agent File Locations

Agents can be placed in:

  • agents/ — The main agents directory (with subdirectories)
  • .aether/agents/ — Project-local agents
  • .github/agents/ — GitHub-integrated agents
Tip: Run aether init after adding new agent files to discover and register them automatically.

Routing & Resolution

When a task arrives, the AgentRouter uses 6 strategies to find the best agent:

  1. Capability match — keyword matching against capabilities array
  2. Section match — domain bucket filtering
  3. RAG semantic search — vector similarity across agent descriptions
  4. File ownership — file patterns mapped to specific agents
  5. Escalation chain — walk escalation targets upward
  6. Random fallback — pick any available agent
Confidence threshold: The router requires a minimum confidence score (default 0.6) before assigning a task. Below that threshold, it falls through to the next strategy. Configure via routing.confidenceThreshold in settings.

Escalation Flow

When an agent fails a task, it escalates upward through the hierarchy:

Worker fails Manager (escalationTarget) Master (cortex-0) ⚡ Circuit breaker

The circuit breaker at the master tier prevents low-priority noise from overwhelming the most expensive model. Each tier has its own failure budget.

Circuit breaker: When the escalation threshold (default 3 failures within 5 min) is hit, further escalations are blocked until the window resets. Configure via escalation.threshold and escalation.windowMs.

What's Next?