The model is the engine. This is the machine.
Abstract
Claude Code, Codex, Gemini CLI — the magic isn’t just the model. It’s the harness: the system that decides what the model sees, what it can act on, what it remembers, and how it recovers when things fail.
This specification takes that machine apart, numbers every part, and says plainly where each one stops working. Four pillars, 29 sheets, four maps you can open and inspect.
Claims
1The Harness
A system disposed around a language model, comprising eleven primitives — each one existing because the previous one ran out of road.
- Instructions
- Context Delivery
- Context Mgmt
- Tools
- Exec Env
- Durable State
- Orchestration
- Sub-agents
- Skills
- Verify + Observe
- Evolution
Memory
The system of claim 1, wherein the model itself remembers nothing: four memory types and an assembly engine decide what reaches the window.
- Working
- Episodic
- Semantic
- Procedural
- Assembly
- Forgetting
Agentic SDLC
The system of claim 2, operated at team scale, wherein agents join the lifecycle and a factory floor replaces the one-way conveyor.
- Evolution
- The 20%
- ADLC Loop
- Ownership
- Factory Floor
- Context Layer
Forward Deployed
The system of claim 3, sold — wherein a deeply technical platform reaches a non-technical buyer, and the chain closes where it began.
- The 2×2
- Outcomes
- ACV Gap
- Primitives
- Two Questions
- Agentic Shift
Something went wrong. Which layer ran out of road?
Pick a symptom. The primitive that owns it lights up on the sheet — and it is almost never the model.
Reported fault
Which layer ran out of road?
When an agent fails, don’t ask “was the model good enough?” — pick a fault and see which harness primitive was missing.