Introducing MIR: An Open Standard for Classifying AI Agent Actions

The Problem: No Common Language

Ask five engineering teams how they describe what their AI agents do, and you will get five different answers. One team calls it "deployment." Another calls it "release." A third uses "push to production." They all mean the same thing, but their monitoring, auditing, and governance systems treat them as unrelated events.

This lack of a common vocabulary is not just an inconvenience. It is a governance gap. If you cannot consistently classify what AI agents are doing, you cannot consistently govern it. You cannot write policies that work across systems. You cannot aggregate audit data meaningfully. You cannot compare risk across domains.

The AI agent ecosystem needed a shared taxonomy. So we built one and open-sourced it.

What MIR Is

The Intended Intent Reference -- MIR -- is an open taxonomy for classifying AI agent actions. It defines 14 enterprise domains, each containing categories that describe the specific actions AI agents perform in that domain.

The domains span the full range of enterprise operations:

• MIR-100 Software Development: deploy, merge, release, rollback, configure
• MIR-200 Security Operations: access grant, incident response, vulnerability scan
• MIR-300 Infrastructure: provision, scale, terminate, configure network
• MIR-400 Data Operations: migrate, transform, backup, restore
• MIR-500 Financial Operations: payment approval, budget change, invoice processing
• MIR-600 HR Operations: onboard, offboard, access review, role change
• MIR-700 Customer Operations: ticket escalation, account modification, SLA review
• MIR-800 Legal and Compliance: contract review, policy update, audit response
• MIR-900 Sales Operations: deal approval, discount authorization, territory change
• MIR-1000 Marketing Operations: campaign launch, budget allocation, content publish
• MIR-1100 Supply Chain: order approval, vendor change, logistics update
• MIR-1200 Research and Development: experiment launch, resource allocation, IP filing
• MIR-1300 IT Operations: service change, incident management, capacity planning
• MIR-1400 Executive Operations: board reporting, strategic decision, M&A workflow

Each domain includes standardized category identifiers, human-readable descriptions, typical risk levels, and common escalation triggers.

How It Works

MIR is distributed as a package you can install in any project:

npm install @intended/open-intent-layer

Using it is straightforward:

import { classify, MIR } from "@intended/open-intent-layer";

// Classify an action description
const result = classify("deploy frontend to production");
// => { domain: "MIR-100", category: "deploy", confidence: 0.94 }

// Reference domains directly
const domain = MIR.domains["MIR-100"];
// => { name: "Software Development", categories: [...] }

// Check if an action matches a category
const isDeployment = MIR.match("deploy frontend to production", "MIR-100.deploy");
// => true

The taxonomy is also available as a JSON file for non-JavaScript ecosystems. Python and Go packages are on the roadmap.

Why We Open-Sourced It

Open-sourcing MIR was a strategic decision, not an altruistic one. We studied the playbooks of companies that built categories by open-sourcing foundational technology.

Databricks did not build the big data market by keeping Spark proprietary. They open-sourced the compute engine and built a commercial platform on top of it. Spark became the standard for distributed data processing, and Databricks became the company that enterprises turned to when they needed a managed, enterprise-grade version.

Confluent followed the same pattern with Kafka. The open-source project defined how the industry thinks about event streaming. The commercial platform handles the operations, security, and governance that enterprises need in production.

HashiCorp did it with Terraform. The open-source tool became the standard for infrastructure as code. The commercial platform -- Terraform Cloud and Enterprise -- added collaboration, policy enforcement, and governance.

MIR is our Spark, our Kafka, our Terraform. It is the foundational layer that defines how the industry classifies AI agent actions. It is free, open, and Apache 2.0 licensed. Anyone can use it, extend it, and contribute to it.

The commercial value is in what you do after classification. Once you know that an AI agent is performing a "production deployment" in the "Software Development" domain, you need to evaluate that against policies, score the risk, issue an authority token, and record the decision in a tamper-proof audit chain. That is Intended -- the Authority Engine, Domain Intelligence Packs, and Audit Chain that turn classification into governance.

The Flywheel

The economics work like this:

• Developers adopt MIR because it solves a real problem -- consistent classification of AI agent actions
• As MIR adoption grows, it becomes the standard vocabulary for AI agent governance
• Teams that classify actions with MIR eventually need to evaluate those actions against policies
• That evaluation -- with risk scoring, domain intelligence, and cryptographic proof -- is Intended

Every team that adopts MIR is a team that might need Intended. Not because we lock them in, but because classification naturally creates demand for evaluation. If you know what your AI agents are doing, the next question is whether they should be doing it.

How to Adopt MIR

For individual developers

Install the package and start classifying. Use MIR categories in your logging, monitoring, and alerting. Even without Intended, consistent classification makes your AI agent operations more observable and auditable.

For platform teams

Standardize on MIR across your AI agent fleet. Use the taxonomy as the common vocabulary for policies, alerts, and dashboards. When every team describes actions the same way, governance becomes possible at scale.

For the community

MIR is on GitHub at github.com/meritt-ai. We accept contributions for new categories, domain extensions, and language-specific packages. The taxonomy is versioned and follows semantic versioning -- breaking changes only in major releases.

What Comes Next

MIR 1.0 covers 14 domains with 80+ categories. We are working on extensions based on community feedback:

• Industry-specific sub-domains (healthcare, financial services, government)
• Composite actions (multi-step workflows that span domains)
• Risk metadata (default risk levels and escalation triggers per category)
• Localization (translations for non-English governance contexts)

The goal is for MIR to become the standard way the industry talks about AI agent actions -- the way HTTP status codes standardized web responses, or the way syslog levels standardized logging severity.

Star us on GitHub, install the package, and start classifying. The AI agent era needs a common language. MIR is that language.