ProvenanceCode Standard · Apache 2.0

Decision-driven development for AI agents.

A structured standard for capturing, enforcing, and auditing AI agent decisions — at the code layer, at runtime, and everywhere between. Open source. Deterministic. Governance without overhead.

Read the Standard → Agent OS docs

AI agents are fast. Their decisions are invisible.

AI coding agents have increased development velocity. They have also increased untraceable architectural decisions, undocumented AI influence, review ambiguity, and governance gaps. Most engineering teams cannot answer: why does this code exist? Was AI involved? Who approved it? What risk was acknowledged?

The same problem applies to production AI agents. They act quickly, but without a structured record of what was decided, why, and by whom, accountability is impossible after the fact.

Why does this code exist?

🤖

Was AI involved?

Who approved the decision?

⚖️

What risk was acknowledged?

Velocity without accountability is fragile.

Four layers. One consistent audit chain.

📄

Decision Records

Structured Decision Evidence Objects (DEOs) are created for every AI-assisted change — capturing intent, risk, and the agent or human responsible. Stored in the repository alongside the code.

🔌

Enforcement Presets

Three levels — Light, Standard, Regulated. Configure governance depth per team or per environment. Rules are YAML. Evaluation is deterministic.

🔗

Merge-Time Validation

Pull request hooks enforce the standard at merge time. Missing decisions fail validation. Unapproved decisions block merge. The gate is configurable.

Human-in-the-Loop

AI may propose. Humans approve. Every decision — made by agent or engineer — is recorded, signed, and attributable. The record does not depend on memory.

The Standard governs code changes.
Agent OS governs runtime execution.

ProvenanceCode Standard operates at the pull request layer — AI coding agents (GitHub Actions, GitHub App, Cursor, Claude Code) submit DEOs with every change. Agent OS operates at the execution layer — any AI agent acting in production is intercepted, evaluated against policy, and recorded before it can touch infrastructure.

Together they form a complete provenance chain: from the first AI-assisted line of code, through review and merge, to every agent action in production. Decision Evidence Objects and runtime provenance records share the same schema. One audit trail, end to end.

AI Coding Agent
Generates code + Decision Evidence Object
Cursor · Claude Code · Copilot · GitHub Actions
ProvenanceCode Standard
Merge-time validation · DEO enforcement
Light / Standard / Regulated preset
Code in production
Your AI agents execute in production
OpenAI · Amazon Bedrock · OpenClaw
Agent OS
Runtime intercept · Policy check · Provenance record
Every action signed before it executes

Structured artifacts. Stored in the repo. Linked to code.

A Decision Evidence Object (DEO) is the structured record created whenever an AI agent or engineer makes a consequential architectural or implementation decision. It captures intent, risk acknowledgement, alternatives considered, and who approved it.

DEOs are stored in the repository alongside the code they describe. They are versioned, diff-able, and auditable. They do not replace code review — they make code review defensible.

v2.0 format
DEC-PROJ-CORE-000001 v1.0: DEC-000001 (also supported)
🎯

Monorepo support

Map DEOs to specific workspaces and sub-projects within monorepos.

🔗

Jira integration

Automatic project and component linking for teams using Jira as a source of record.

📁

Hierarchical structure

Project / subproject / component — DEOs mirror your engineering org structure.

Backward compatible

v1.0 format fully supported. Migration to v2.0 is optional and incremental.

Open standard. Enterprise governance.

The ProvenanceCode Standard is Apache 2.0. Read the specification, implement the GitHub Action, or deploy Agent OS for runtime governance — pick the layer that fits where you are today.

Read the Standard Book a call