Observe
Pull the real signal
Events, files, metrics, and conversations enter the system.
01 The Agentic OS Field Guide
An AI operating system is not a perfect mission control. It is a living stack of context, constraints, workflows, roles, and connections that keeps getting better at one domain.
Click the globe to reveal its layers
One architecture · Three scales · Any model harness
02 The operating loop
A folder becomes an operating system when evidence can travel through a repeatable, verifiable cycle.
Observe
Events, files, metrics, and conversations enter the system.
Decide
Purpose and constraints narrow the valid course of action.
Execute
An agent chooses the smallest proven workflow that fits.
Verify
Hooks test permissions, evidence, quality, and safety.
Learn
Decisions and results strengthen the substrate for next time.
03 Six layers, one architecture
The layer names stay stable. The application changes with the scale of the system.
Soul, purpose, point of view, pointers
The canonical definition of the system: who it serves, what it optimizes for, how it reasons, and where deeper context lives.
“You are a fractional CFO for a bootstrapped SaaS. Optimize for cash and clarity.”
identity.mdprinciples.mdcontext-map.md04 The portability layer
Your OS should not live inside one vendor file. Keep the identity and workflows neutral. Treat every harness file as an adapter.
identity.md + rules/ + skills/One truth, many interfacesCLAUDE.mdA provider adapter that points Claude at the canonical identity, skills, checks, and working context.
AGENTS.mdDurable repository guidance for Codex. Nested files can specialize instructions without forking the soul.
GEMINI.mdA harness-specific map to the same neutral source, tools, and verification contract.
Principle Provider files point inward. They do not become the soul.
05 Context rot simulator
Drag time forward. Every layer changes at a different rate, so maintenance must be designed—not remembered.
Tools and active agents already need a pulse check.
06 Build your V0
Build the smallest complete loop. Earn complexity by using it.
Create a wiki folder. Add the evidence you already trust: docs, transcripts, exports, decisions.
Write one neutral identity.md: purpose, audience, point of view, boundaries, and context pointers.
Point CLAUDE.md, AGENTS.md, or your chosen harness file to the same canonical identity.
Choose painful workflows you have repeated. Capture inputs, steps, judgment points, and definition of done.
Only if the two skills form a real job. Give the agent ownership, limits, and a handoff contract.
Every generic answer becomes a rule. Every risky action becomes a hook. Every stale connector gets a review date.