System Overview
AI Delivery OS is a three-layer operating system for AI-assisted software delivery. It governs how AI agents plan, implement, and validate code changes by providing structured evidence from the codebase, constraining agent decisions based on that evidence, and recording every decision for debugging and calibration.
The system is not an autonomous agent. It is a decision-support tool that reduces the probability of wrong plans by giving AI agents grounded, reliability-scored context — and constraining them to operate within evidence-backed boundaries.
Three Layers
Knowledge Engine
Builds structured evidence from the codebase. Parses code, indexes modules, traces dependencies, detects patterns, and produces a reliability-scored Execution Context Object (ECO) for every task.
Task Orchestrator
Turns requirements into bounded execution slices. Classifies work into lanes, selects strategies, decomposes tasks, and assigns scoped boundaries to each implementer agent. Constrained by the ECO.
Decision Ledger
Records why the system believed something, chose something, and did something. A unified trace schema spans all stages. Operational contracts enable continuous improvement.
Dual Inputs
Every task has two inputs of equal weight:
| Input | What it provides | Authority |
|---|---|---|
| Spec file | Intent · scope · acceptance criteria · risk notes | What the developer wants |
| Codebase | Modules · dependencies · graph edges · patterns · gate results | What actually exists |
The knowledge layer's job is to resolve the spec's intent against the codebase's reality. The Execution Context Object (ECO) is the output of that collision.
End-to-End Flow
Key Numbers
What the System Does Not Do
The knowledge layer does not understand code. It retrieves structural facts, measures its own reliability, and communicates its limits. The reasoning happens in the LLM that consumes the ECO — the knowledge layer provides the evidence.
See Design Boundaries for the full list of known limitations and gaps.