Timeline: March 3 → March 16. ~13 active session days. 23 handoff docs. 251 tests. $12/month server. Half a penny per product.

Two and a half weeks ago i opened my laptop in a hospital room and started experimenting with AI agents. i was trying to automate a small task. Instead i ended up building an orchestration system, a publishing pipeline, and several live projects — all running on a low-cost Linux VPS. No Kubernetes. No cluster. No expensive AI infrastructure. Just a small Linux box and a lot of software… although i experimented with several platforms before landing here, this has been the sweet spot.

As i started building, this phrase, this AI created philosophy started repeating from GPT:

Agents generate ideas. Software controls reality.

It’s been 2.5 weeks since i opened my laptop in a hospital room and started learning how to build with AI agents.

Here’s what exists.

The numbers

  • $0.005 per product through the full pipeline
  • 62+ products generated and published
  • 251 tests written
  • 23 handoff documents across 11 session days
  • $12/month server
  • Timeline: 2.5 weeks

The Numbers That Surprised Even Me

  • 33 products via v2 pipeline + ~17 products via dominate-orchestrator + a 12-product viral batch
  • First live publish milestone: March 15
  • $0.005 per product through the full pipeline — half a penny in AI cost per product
  • 251 tests written (settled at 213)
  • 23 handoff documents across 11 session days
  • 1,273 Printify blueprints cached and searchable
  • 6 real fonts, 5 design recipes, 10 product templates — premium typography built in
  • 26 Gmail labels across 18 companies with automated routing
  • Tax spreadsheet — Google Sheets integration for expense tracking
  • AI Execution Engine — shipped as a $29 product with 18-page guide, 6 architecture modules, and AI copilot workflow
  • HashiCorp Vault — real secret management with scoped tokens, not just env files
  • 36% malicious rating found on skillsmp.com — AMOS stealer distribution identified and flagged
  • **Installed LLM Qwen 7B on llama-server

When you zoom out, the system that emerged looks like this:

Execution Control

  • Orchestrator
  • State machine
  • Event-driven execution
  • Closed action vocabulary

Safety

  • Validation layer
  • Budget gates
  • Token controls
  • Human approval gates
  • Price band enforcement
  • Transparency enforcement (corner-pixel sampling)
  • Deterministic validators
  • Product binding invariant

Memory

  • Shared state JSON
  • Layered memory (3 tiers with topic-routed loading)
  • Research directories
  • Handoff documents
  • Session logs
  • Standing rules
  • Prompt caching (stable prefix + dynamic suffix)

Reliability

  • Idempotency (SHA-256 hash)
  • Job ledger
  • Model routing
  • State machine fixes
  • Retry loop (POLISH→REVALIDATE up to 5 cycles)

Cost Optimization

  • Subagent model routing (haiku/sonnet/opus tiered)
  • OpenRouter lowest-cost routing
  • $0.005 per product through full pipeline
  • Token limits per stage

The Philosophy

Through all of this, one idea kept coming back:

Agents generate ideas. Software controls reality.

The model should never control the system. The system must control the model.

The system that emerged from all of this became the AI Execution Engine — a blueprint for anyone who wants to build their own.

More at aitinkers.fun