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:
- Scrape — agents pull relevant content from defined sources
- Score — relevance and quality scoring against defined criteria
- Filter — candidates that pass scoring thresholds advance
- Generate — content production from vetted source material
- Sanitize — OPSEC review ensures no operational details leak into public content
- Review — human approval gate before anything publishes
- 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