Skip to content

What the Agents Produce

The fleet isn't an experiment. It produces operational output across multiple domains.

Content Pipeline

Automated content production with human quality gates:

  1. Scrape — agents pull relevant content from defined sources
  2. Score — relevance and quality scoring against defined criteria
  3. Filter — candidates that pass scoring thresholds advance
  4. Generate — content production from vetted source material
  5. Sanitize — OPSEC review ensures no operational details leak into public content
  6. Review — human approval gate before anything publishes
  7. Publish — automated deployment to target platforms

The pipeline handles the volume work. The human handles the judgment calls.

Code Arena

Frontier LLMs compete head-to-head on identical specifications:

  • Identical prompts sent to multiple models simultaneously
  • Side-by-side deployment of results for comparison
  • Structured evaluation — functionality, code quality, adherence to spec, edge case handling
  • Scoring and ranking that feeds back into dispatch routing decisions

This produces empirical data on model performance for specific task types, replacing opinions with measurements. When the dispatch system routes a coding task to a particular model, that routing is based on arena results, not vibes.

Trading Analytics

Multi-stage analytical pipeline for market analysis:

  • Deterministic technical analysis — multi-timeframe TA with defined indicators and invalidation criteria
  • LLM invalidation layer — AI models used as red team, actively trying to kill trade setups rather than confirm them
  • Risk management — position sizing, exposure limits, defined exit criteria
  • Quantitative gating — no trade proceeds without meeting minimum statistical thresholds

Philosophy

LLMs generate narrative confidence, not calibrated probabilities. The architecture uses them for what they're good at (pattern recognition, red-teaming setups) and keeps them away from what they're bad at (price prediction, signal generation).

Fleet Observability

The fleet monitors itself:

  • Host-level telemetry — CPU, memory, disk, network across all environments
  • Model-level metrics — token usage, cost tracking, response times, error rates
  • Real-time visualization — multiple dashboard views with different focus areas
  • Compressed telemetry — efficient data representation for long time horizons
  • Observer independence — monitoring systems are designed so their observation doesn't alter the system being observed

Parametric 3D Modeling

Vision models produce editable parametric CAD output from reference images:

  • Input — photographs, sketches, reference images of physical objects
  • Processing — vision model analysis of geometry, dimensions, relationships
  • Output — parametric CAD files (not mesh — actual editable geometry with constraints)
  • Application — rapid prototyping, reverse engineering, documentation of physical assets

Scale

  • 100+ project planning documents in the knowledge vault
  • Multiple frontier models in active rotation
  • Automated content production at consistent velocity
  • Multi-domain output (content, code, analytics, security, 3D) from a unified fleet
  • Everything behind Cloudflare Zero Trust access controls