Docs / Agent Fleet

Agent Fleet

AI OS orchestrates 51 specialized sub-agents powered by Claude Opus 4.8 with effort-based routing. Instead of separate models, agents are assigned effort levels (xhigh, high, low) that control reasoning depth. Each agent has a defined role, capabilities, and escalation path managed through Markdown files with YAML frontmatter.

Strategic Tier — Opus 4.8 (effort: xhigh)

Maximum reasoning depth for decisions requiring deep analysis, architectural judgment, and cross-system awareness. Uses adaptive thinking for complex multi-step planning.

AgentRole
OrchestratorCentral intelligence. Decomposes complex requests, delegates to sub-agents, manages inter-agent data flow, synthesizes final outputs.
ArchitectSystem design and technical architecture. Evaluates trade-offs, defines data models, plans migrations.
ReviewerCode and content review. Identifies quality issues, enforces standards, provides actionable improvement suggestions.
Security AuditorVulnerability assessment, dependency auditing, CSP review, authentication flow analysis.

Professional Tier — Opus 4.8 (effort: high)

The workhorse tier handles the majority of tasks with balanced quality and speed. Opus 4.8 at high effort provides the same model quality with optimized thinking depth.

AgentRole
ResearcherDeep research with source synthesis, citation tracking, and structured output generation.
CoderFull-stack development, debugging, refactoring, and test authoring across languages.
WriterLong-form content, copywriting, documentation, and tone-adaptive writing.
Design SystemDESIGN.md creation, brand reasoning, color theory, WCAG compliance, component specifications.
Media ProducerRemotion video generation, Google Vids integration, asset composition.
Lead GenProspect scraping, enrichment, scoring, and personalized outreach sequence generation.
Marketing HubMulti-platform content pipelines, scheduling, A/B variant generation, performance tracking.
Product FactoryDigital product creation, marketplace listing, Etsy/Gumroad publishing automation.
Vibe DesignerPrompt-driven UI generation with predictive heat maps and interaction flow analysis.
Blender 3DText-to-3D environments and product renders via Blender MCP integration.
Knowledge GraphAuto-categorizing knowledge ingestion, semantic linking, and radial graph visualization.
Golden LoopGemini Gems synced to NotebookLM notebooks for continuous knowledge amplification.
SynthesisCross-agent output merging, conflict resolution, and cohesive deliverable assembly.
AutomatorCRON-scheduled routines, workflow triggers, and conditional automation chains.
Browser AgentHeadless browser control for scraping, form filling, and multi-step web interactions.
Report CompilerStructured report generation from multiple data sources with charts and tables.
Research ArchitectResearch project planning, methodology design, and literature review structuring.
Data WranglerData cleaning, transformation, format conversion, and statistical analysis.
QATest plan creation, regression testing, edge case identification, and quality metrics.

Scout Tier — Opus 4.8 (effort: low)

Fast agents for quick lookups and lightweight tasks. Low effort minimizes thinking time while retaining Opus 4.8's full capability when needed.

AgentRole
ScoutQuick fact-checking, simple lookups, and rapid triage of incoming requests.
Social IntelSocial media monitoring, sentiment analysis, and trend detection.
Routine RunnerExecutes pre-defined CRON routines with minimal overhead.

Economy Tier — DeepSeek V4

Maximum throughput at minimal cost for batch operations.

AgentRole
DeepSeek WorkerBulk content generation, template expansion, and high-volume text processing.
Batch RunnerQueue-based mass production with automatic retry and output routing.

Realtime Tier — Grok-3

AgentRole
Grok RealtimeLive data queries, current events, real-time web intelligence, and social platform analysis.

Agent Definitions

Each agent is defined as a Markdown file in .claude/agents/ with YAML frontmatter specifying its metadata:

---
name: researcher
model: opus-4.8
effort: high
tier: professional
escalates_to: orchestrator
tools:
  - web_search
  - vault_read
  - vault_write
---

# Researcher Agent

You are a deep research specialist. Your role is to...

Frontmatter Fields

  • name — Unique identifier used in API calls and routing
  • model — AI model to use (opus-4.8, gemini-omni-flash, deepseek, grok)
  • effort — Opus 4.8 effort level: xhigh (strategic), high (professional), low (scout)
  • tier — Cost/capability tier for routing decisions
  • escalates_to — Agent to escalate to when confidence is low
  • tools — List of tools/capabilities available to this agent

Creative Tier — Gemini Omni

Multimodal creative generation powered by Gemini Omni Flash. Produces video, images, audio, thumbnails, and social clips from text prompts.

AgentRole
Media Producer (Muse)Creative pipeline orchestration, production strategy, brand consistency across media.
Vibe Designer (Pixel)Prompt-driven UI generation, predictive heat maps, interaction flow analysis.
Video Creator (Reel)Text/image/audio to video with physics simulation. Up to 1080p, 60s.
Audio Producer (Sonance)Voiceovers, music, sound effects, podcast audio from text prompts.
Thumbnail GenPlatform-optimized thumbnails with A/B variants for YouTube, social, blog.

SEO Agency Sub-Agents

Five specialized agents that run in parallel during SEO audits, each analyzing a different dimension of the target domain.

AgentRole
SEO KeywordKeyword research, gap analysis, cannibalization detection, local targeting.
SEO TechnicalCrawl analysis, Core Web Vitals, HTTP status codes, security, structured data.
SEO CompetitorTop 10 competitor profiling, content velocity, schema adoption, ranking overlap.
SEO ContentContent inventory, thin content detection, topic clusters, meta data quality.
SEO BacklinkReferring domains, toxicity detection, broken link recovery, anchor text analysis.

Legal Department

Handles licensing agreements, compliance, contracts, and intellectual property protection.

AgentRole
General Counsel (Justice)Chief Legal Officer. License agreements, IP protection, regulatory compliance, dispute resolution.
Compliance Officer (Shield)GDPR/CCPA enforcement, audit trails, policy monitoring, API usage compliance.
Licensing Attorney (Covenant)Software license agreements, white-label terms, SaaS licensing, usage rights.
Contract Specialist (Clause)Contract generation, review, lifecycle management, template library.

Persistent Tier — Hermes MCP

The always-on background worker tier. Hermes runs 24/7 on the platform and handles tasks that require persistence beyond the active session.

AgentRole
Hermes DelegateBackground task execution with progress pings. Walkaway mode for long refactoring tasks with mobile replies.
Hermes Cron24/7 scheduled jobs: inbox summaries, news briefs, GitHub backups, comment monitoring, health checks.
Hermes Approval GatePauses destructive actions and pushes approval prompts to Telegram/Slack/Discord for mobile review.
Learn more: See the Hermes Agent documentation for setup, API endpoints, and walkaway mode details.

Escalation Paths

Agents can escalate tasks upward when they encounter requests beyond their capability. The escalation chain follows the tier hierarchy:

  • Economy → Scout → Professional → Strategic (Orchestrator)
  • The Orchestrator is the terminal escalation point and can re-delegate to any agent