Agentic AI Discipline Protocol (AADP)

A governance framework for keeping agentic-AI work structured, auditable and anchored to explicit human decisions.

Why discipline matters

Agentic systems can act across many steps and tools. Without clear constraints, evidence and decision boundaries, automation can drift away from the owner's intent. AADP addresses that governance problem.

Core themes

Explicit decisions

Important requirements and owner choices are made visible rather than left to inference.

Invariants

Critical rules remain stable across tasks, agents and implementation cycles.

Auditability

Evidence and verification make it possible to examine what was done and why.

Anti-drift controls

Defined boundaries help prevent automated work from silently changing scope or policy.

AADP in plain language

Agentic AI Discipline Protocol is a written working agreement between AI agents and a human decision-maker. It specifies who may decide, what the agent may change, what evidence supports a claim and what to do when uncertainty or a new problem appears.

For example, an agent asked to improve a materials workflow should work within the agreed task, verify the records it cites, report failed checks and seek the required human decision before consequential action. AADP provides a discipline for that work; the application must still implement and test its own controls.

How the protocol is organised

Role separation: distinguish orchestration, execution and human decision authority.
Eleven invariants: rules for claim evidence, verified references, scope, rechecking and source-backed changes.
Versioned context: retain decisions, session state and change history across sessions.
Failure escalation: record deviations and strengthen documented controls when problems recur.

From AADP to AADP-Ops

AADP is the published protocol. AADP-Ops explores its operational application. The mobile materials demonstrator illustrates controlled issue, separate approval, quarantine and traceable outcomes using synthetic records. It does not yet perform live AI part identification or connect to AMOS.

Explore the materials case study

Read the published source

Authored by Vijayaprakash Govindarajan and published under the MIT licence. The repository contains the protocol, prompt and report skeletons, context templates and an author-reported MAYA origin case study. That bounded case does not establish performance across all deployments or guarantee elimination of AI errors.

Read AADP on GitHub

Current status

Published framework. AADP is part of LAKSAHA's work on practical governance for AI-assisted systems and operations.