Methodology · MBSE & AI Agent

Two layers: MBSE method profiles (MagicGrid, OOSEM, Harmony-SE) for knowledge slices; ADPS pattern language (cognitive function × execution topology) for Agent experiment reviews.

MBSE methods set modeling rhythm; ADPS vocabulary frames Agent experiment notes.

MBSE methodology packs

These packs drive retrieval filters, diagram pacing, and review prompts inside the engineering knowledge base and AI4MBSE. In a live session you can switch packs — next-diagram suggestions and knowledge focus follow the active pack.

MagicGrid

Problem · Solution · Implementation

Common in rail, energy, and complex infrastructure. Typical next views after structure: activity and sequence.

Cases: PSD · wind farm

OOSEM

Operational ↔ engineering alignment

Strong for regulated domains. Typical path: requirements and use cases before structure and verification clues.

Cases: mine positioning · infusion pump

Harmony-SE

Use-case driven

Strong for embedded and robotics. Typical path: scenarios and interactions, then states and internal structure.

Cases: robot vacuum · AGV

Nine common views · one at a time

The experiment paces familiar SysML views — package, block definition, internal block, use case, requirements, activity, state machine, sequence, and parametric — without claiming every project must use all nine. The active methodology pack only changes which view is suggested next.

  • PKG
  • BDD
  • IBD
  • UC
  • REQ
  • ACT
  • STM
  • SD
  • PAR

Interactive sketch (methodology switch + next chips): AI4MBSE local demos.

Locale-aware method names

When the host UI locale is Chinese or English, methodology display names and next-step prompts follow the chrome language. Modeling dialogue may still mix languages — the pack itself is engineering rhythm, not a translation gimmick.

AI Agent methodology · ADPS patterns

As Agents move from “models that answer” to systems that join workflows, engineering focus shifts to context, evidence, state, tools, permissions, collaboration, and audit. This personal lab uses the public ADPS pattern catalog as vocabulary for experiment design reviews — not a product certification.

Two-axis framework

ADPS organizes patterns on two axes. The vertical axis is cognitive function — what the Agent is doing. The horizontal axis is execution topology — how data and control flow.

Cognitive functionOne-line definition
PerceptionWhat the Agent can see now, and in what form
MemoryWhich experience, evidence, and state survive across turns
ReasoningCompile inputs, evidence, and state into executable judgment
ActionGround judgment in tools, APIs, files, workflows, and the real world
ReflectionEvaluate, revise, and retain outputs, failures, and trajectories
CollaborationDivide, parallelize, review, and hand off among Agents
GovernanceBound permissions, risk, audit, and accountability
TopologyVerbFits when
ChainPassStable steps; prior output feeds the next
RouteChooseClassify first, then pick model, tool, flow, or human review
ParallelFan outRun branches together; trade cost for quality or latency
OrchestrateCoordinateA coordinator owns goals, task ledger, and aggregation
LoopIterateGenerate, observe, revise, until converge or trip a fuse
HierarchyLayerLayered duty, permission, memory, or protection boundaries

Core pattern matrix (v0.3)

ADPS v0.3 catalogs 28 core patterns, 5 extensions, and 1 candidate (34 specs). Empty matrix cells are intentional — they mean no stable first-order structure yet, or an existing pattern already covers the main problem.

Function ChainRouteParallelOrchestrateLoopHierarchy
Perception P2P1P4P3
Memory M2M3M2 / M4M1
Reasoning R1R2R3R4
Action A3A1A2A4
Reflection F1F2F4F3
Collaboration C4C2C3C1
Governance G3G1G4G2
  • Perception — P1 Context Triage · P2 Semantic Compaction · P3 Progressive Discovery · P4 Multi-Modal Fusion
  • Memory — M1 Hierarchical Retention · M2 RAG · M3 Progress Tracking · M4 Failure Diary · M5 Procedural Memory (ext.)
  • Reasoning — R1 Chain of Thought · R2 Complexity Routing · R3 Parallel Exploration · R4 Iterative Hypothesis · R5 Talker-Reasoner (ext.)
  • Action — A1 Tool Dispatch · A2 Plan-Execute · A3 Prompt Chaining · A4 Guardrail Sandwich · A5 Minimal Tool Set (ext.)
  • Reflection — F1 Generate-Critique · F2 Skill Pack · F3 Experience Replay · F4 Self-Heal Loop
  • Collaboration — C1 Hierarchical Delegation · C2 Fan-out Aggregate · C3 Adversarial Review · C4 Handoff Chain · C5 Sub-Agent Isolation (ext.) · C6 Choreography (candidate)
  • Governance — G1 Approval Gate · G2 Blast-radius Control · G3 Progressive Commitment · G4 Observability · G5 Hooks Pipeline (ext.)

Five engineering contracts

v0.3 recommends that production Agents make these contracts explicit — the main control surface between model judgment and engineering state:

  1. Context — what enters context, what is a handle, what is deferred or dropped (Perception).
  2. Evidence — source, version, scope, citation, provenance, trace (Memory / RAG).
  3. Decision — how model output becomes structured Route / Reasoning / Plan decisions (Reasoning).
  4. Action — tool metadata, parameter provenance, permissions, approval, human review, rollback (Action · Governance).
  5. Trace — LLM calls, tool calls, handoffs, guardrails, approvals, and state updates under one trace id (Governance).

Source & lab stance

This section summarizes the public ADPS white paper Pattern catalog & selection framework (v0.3). It offers architecture vocabulary for experiment reviews — not certification of a product. This personal lab treats ADPS as shared pattern language for Agent technical exploration, and is glad to learn from / contribute notes as a community participant. Full specs follow the ADPS source.

Related write-up: personal AI4MBSE simulation experiment report (author’s peer technical-experiment site — not an enterprise product case).

Peer research link

How the AI4MBSE experiment uses both layers

MBSE packs shape retrieval rhythm and next-diagram hints; ADPS vocabulary frames Agent experiment review (route intent, human approval gates, evidence-aware memory). Details: AI4MBSE and KB notes · Disclaimer