Agentic Case Management Notation
An open standard extending CMMN with native primitives for AI agent orchestration.
v1.0 Working Draft April 2026
Overview
ACMN is a notation for modelling case management processes that include autonomous AI agents as first-class participants alongside human actors. It extends the Case Management Model and Notation (CMMN) 1.1 standard published by the OMG, adding elements for agent orchestration, tool integration, confidence-driven progression, domain governance, and autonomous case creation from external signals.
ACMN preserves all CMMN concepts and terminology — cases, stages, sentries, milestones, case files, human tasks, and discretionary items — and adds new elements where CMMN provides no equivalent. The standard is implementation-agnostic: it defines what elements exist and how they behave, not how to build a runtime.
Example: Insurance claim case with intake pipeline
Signal source → intake agent → intake queue → case plan model with stages, agent nodes, tool nodes, evaluation, and human escalation.
Four-layer taxonomy
| Layer | Name | Contents |
|---|---|---|
| 1 | CMMN inherited | Case, case plan model, case intake, stage, milestone, sentry, case file, human task, discretionary item, plan item decorators |
| 2 | Agent primitives | Agent node, tool node, guardrail node, evaluation node, handoff node |
| 3 | Context layer | Five wire types, typed ports, domain context (vocabulary, schemas, rules, decision tables) |
| 4 | Runtime semantics | Case and agent lifecycles, sentry evaluation, wire buffering, event bus, state persistence |
Element catalogue
Nine plan item types — four inherited from CMMN and five new — plus the case intake pipeline.
CMMN inherited
- Stage — container with cognitive modes (gather, analyse, draft, review, decide)
- Human task — work performed by a human actor
- Milestone — checkpoint with confidence criteria, snapshots, revocation
- Sentry — guard with confidence, budget, timer, event on-parts
Case intake (ACMN)
- Signal source — email, webhook, file watch, schedule, event, API
- Intake agent — classifies signals, extracts data, assesses confidence
- Intake queue — review, deduplication, routing, batching
ACMN extensions
- Agent node — autonomous AI participant with 9 property groups (persona, model config, reasoning strategy, confidence, ports, etc.)
- Tool node — external capability with input/output schemas and invocation policies
- Guardrail node — binary constraint checking (rule-based, LLM-based, hybrid)
- Evaluation node — quality assessment with criteria, feedback loops, retry limits, 4 evaluator types
- Handoff node — structured context transfer with reasoning traces
What ACMN adds to CMMN
| Capability | Description |
|---|---|
| Typed data wires | Five wire types (data, confidence-gated, escalation, event, case file) with port-based connections, buffering strategies, and inline transforms. |
| Confidence model | Agents track self-assessed confidence with trajectory analysis. Milestones can require confidence thresholds. Wires gate on confidence scores. |
| Domain context | Shared vocabulary, typed entity schemas, and domain rules that cannot be relaxed. Agents reason with consistent domain language. |
| Decision tables | DMN-compatible decision tables in the domain context. Agents reference them for structured decisions. Evaluators use them as deterministic scoring rubrics. |
| Evaluation loops | Independent quality assessment with structured criteria, weighted scoring, and feedback-retry cycles. Four evaluator types: LLM, agent, human, composite. |
| Governance layers | Three-tier constraint model: domain rules → case-level guardrails → agent-level tool policies. Compliance by design. |
| Milestone revocation | Milestones can be revoked when conclusions are invalidated. Case state rolls back to achievement snapshot. Agents receive revocation context. |
| Case intake | Autonomous case creation from external signals. Signal sources feed an intake agent that classifies, extracts data, deduplicates, and routes to case instantiation. |
Conformance levels
| Level | Name | Scope |
|---|---|---|
| 1 | ACMN Core | Basic case lifecycle, agent nodes, tool nodes, data wires, typed case file. |
| 2 | ACMN Extended | Level 1 + confidence milestones, sentry composition, all wire types, guardrails, evaluation nodes, handoffs, case intake, event bus. |
| 3 | ACMN Complete | Level 2 + domain context (vocabulary, schemas, rules, decision tables), state persistence, cross-case patterns, full visual notation. |
Download
ACMN Standard v1.0 — Working Draft
April 2026 · 1,162 paragraphs · 7 UML class diagrams · 4 visual notation figures
Download specification (.docx) View on GitHubReference implementation
The ACMN Platform is the open-source reference implementation targeting Level 3 conformance. It uses a four-component polyglot architecture:
| Component | Stack | Responsibility |
|---|---|---|
| Visual editor | TypeScript, React, React Flow | Design canvas, property panels, three modes (design/test/live) |
| Comms engine | Node.js, Express, MCP SDK | MCP gateway, webhook adapters, editor API |
| Execution engine | Java, Spring Boot, PostgreSQL | Case lifecycle, sentry engine, wire engine, audit trail |
| Logic engine | Python, FastAPI, LiteLLM | LLM abstraction, context assembly, confidence scoring |
License: Apache 2.0 · Source: github.com/acmn-platform