Browse docs▾
How it works
Coding OS wraps your agent in a disciplined loop and gives it three cognitive layers to draw on. Nothing here is advisory — hooks enforce each step, so the agent can't skip the plan or claim “done” without verifying.
The loop
Every non-trivial request runs through a complexity gate and then five phases. Simple work passes straight through; complex work triggers a deeper cycle.
- Classify
- Orient
- Plan
- Execute
- Verify
Classify
Gate the request by complexity (Cynefin × dimensions) before touching anything.
Orient
Read only what the task needs — graph, docs, memory. Attention is the scarce resource.
Plan
State the approach, the smallest correct change, and the risk — before writing code.
Execute
Implement the minimal diff. Commit each logical unit.
Verify
Run the real verification — not “it should work”. Reading code is not verifying.
The three brains
graph_os — Navigate
Extractors index your code, docs, and configs into a queryable knowledge graph. Instead of grepping and re-reading files, the agent asks the graph directly: who calls this function, what breaks if I change it, what does this rename touch, what is the API surface. One targeted query — a few hundred to a few thousand tokens — replaces reading dozens of files, and the saving grows with the size of the codebase.
thinking_os — Remember
Cross-session memory, learning, and metrics. Sessions record observations, decisions, and failure modes; the next session inherits a digest instead of starting from zero. Patterns that recur get promoted; stale ones decay. This is what turns a stateless agent into one that gets faster at your codebase over time.
board_os — Plan
A Scrumban board with WIP limits, a definition of ready, and lifecycle hooks. Work is expressed as tasks with acceptance criteria the agent must satisfy — a contract, not a vibe. The board is the single source of truth for what is in flight, queued, blocked, or done.
Enforcement, not etiquette
The loop is backed by hooks that fire on the agent's tool calls. They block a code edit that has no traced spec, a commit that violates the message contract, or a “done” without a passing verification. On Claude Code every hook fires; on Codex CLI the Bash-event hooks fire. Discipline the agent can rationalize away is discipline it doesn't have — so Coding OS makes the important steps impossible to skip.
See exactly what each runtime enforces →