Closing the operational knowledge gap in U.S. manufacturing with open-source, UNS-native, cybersecurity-aware multi-agent AI.
📊 See
IMPACT.mdfor adoption signals, citations, and the engagement roadmap.
U.S. manufacturing faces a converging crisis:
- 2.1 million unfilled positions projected by 2030 (Deloitte / Manufacturing Institute)
- 40% of the skilled workforce is within 10 years of retirement, taking 30+ years of institutional knowledge with them
- Small-to-mid manufacturers (SMMs) — the backbone of the CHIPS Act supplier ecosystem — lack the capital, OT expertise, and AI safety literacy to deploy agentic AI without risk
Off-the-shelf AI tools are not designed for the shop floor. They are:
- Not OT-aware (no Purdue zoning, no ISA-95 context)
- Not UNS-aware (cannot speak Sparkplug B or OPC UA natively)
- Not air-gap tolerant (assume cloud connectivity)
- Not CMMC-aligned (non-starter for defense-adjacent suppliers)
- Not auditable (no governance lineage, no human-in-the-loop discipline)
Gartner projects 40%+ of agentic AI projects will be cancelled by 2027 due to unclear value and inadequate risk controls. This framework is the missing reference implementation — UNS-aware, OT-aware, governance-aware — that makes agentic AI safe to deploy in manufacturing.
This framework is designed for U.S. manufacturers being asked to do more with fewer experienced people, under tighter compliance requirements, on shorter timelines.
Small-to-mid manufacturers (SMMs) — approximately 99% of U.S. manufacturers, including Tier-2 and Tier-3 suppliers feeding the CHIPS Act semiconductor ecosystem. SMMs cannot afford enterprise-scale data and AI teams; this framework gives them a reference architecture deployable on a laptop with Ollama, scalable to AWS Bedrock or Azure OpenAI when production-ready.
Defense industrial base subcontractors — need CMMC Level 2 / NIST SP 800-171 alignment as a non-negotiable prerequisite for DoD contracts. The framework's policy engine, Ed25519-signed audit trails, and Purdue-zone enforcement are built to support that posture.
CHIPS Act and reshoring brownfield plants — inherit legacy SCADA, MES, and historian estates. The UNS Context Broker bridges legacy OT and modern AI without rip-and-replace, preserving capital while accelerating Industry 4.0 maturity.
NIST MEP Network centers and Manufacturing USA institutes — need free, open, standards-aligned reference architectures they can recommend to their thousands of SMM clients without endorsing any one vendor.
Academic and research labs — get a reproducible benchmark suite (IABENCH-v1) and a synthetic UNS dataset to advance industrial agentic AI research without negotiating proprietary data access.
The framework is designed to move five concrete needles:
1. Knowledge cliff mitigation. The Tacit-Knowledge Curator captures retiring-expert tribal knowledge into a governed, queryable layer before it leaves the shop floor. A junior technician gets senior-engineer answers on-demand, grounded in cited sources. This addresses the 30+ years of institutional knowledge projected to retire by 2030.
2. AI safety by construction. Purdue zoning, OPA policy gates, Ed25519 audit trails, and human-in-the-loop escalation mean operators retain accountability. This is the missing layer behind the 40%+ agentic-AI project cancellation rate Gartner projects through 2027 — most failures are governance failures, not model failures.
3. Standards-first compliance. ISA-95, OPC UA, MQTT Sparkplug B, NIST SP 800-82, CMMC L2, and NIST AI RMF aren't bolted on — they are architectural primitives. Suppliers can demonstrate compliance posture without expensive retrofitting, lowering the cost of entry into the defense industrial base.
4. Open-source multiplier. As a free reference implementation under MIT license, the framework gives MEP centers, system integrators, and academic groups a starting point they can adapt rather than rebuild. Public infrastructure for industrial AI.
5. CHIPS-era supply chain resilience. Tier-2 and Tier-3 suppliers gain operational visibility and decision intelligence comparable to what their large-customer fabs already have, narrowing the resilience gap that has historically slowed reshoring of high-complexity manufacturing.
Ten specialized agents collaborate in a UNS-mediated group chat, governed end-to-end by an immutable lineage bus:
| Agent | Role |
|---|---|
| Operational Intent | Parses operator NL queries → typed Intent (EN/ES/VI) |
| UNS Context Broker | Mediates ALL tool calls; enforces Purdue zoning; resolves ISA-95 paths |
| Telemetry & Historian | Reads OPC UA, MQTT Sparkplug B, PI, Ignition, Snowflake, Timestream |
| Anomaly & Root-Cause | Multivariate AD (Matrix Profile + Isolation Forest) + FMEA traversal |
| Tacit-Knowledge Curator | RAG over SOPs, expert interviews, OEM manuals, work orders |
| Safety / Guardrail | OPA policy engine; reversibility classifier; CMMC L2 gate |
| Work-Order & MES Dispatch | Writes to CMMS/MES with dry-run diff + idempotency |
| Governance & Lineage | OpenLineage emit + Ed25519 signing + NIST AI RMF mapping |
| HITL Supervisor | Confidence-thresholded human routing (Slack/Teams/email) |
| Shop-Floor Copilot UI | Role-aware Streamlit operator surface |
- Python 3.11+
- Docker + Docker Compose
- (Optional) Ollama for local LLM inference
git clone https://github.com/adris-misra/multi-agentic-framework.git
cd multi-agentic-framework
make install
cp .env.example .env # fill in your API keys
make demo # starts the full docker-compose stackThe demo stack starts:
- MQTT broker at
mqtt://localhost:1883(Eclipse Mosquitto) - ChromaDB at
http://localhost:8000 - Jaeger tracing UI at
http://localhost:16686 - Operator Copilot at
http://localhost:8501
# Install Ollama: https://ollama.com
ollama pull llama3.1:8b
LLM_PROVIDER=ollama make demo
# Then open examples/05_smm_starter_kit.ipynb| Standard | How this framework addresses it |
|---|---|
| ISA-95 / IEC 62264 | UNS topic taxonomy, ISA-95 hierarchy in every AgentMessage |
| MQTT Sparkplug B v3 | Native decode/encode; UNS Context Broker validates all paths |
| OPC UA (IEC 62541) | Async read-only client with zone-enforcement wrapper |
| NIST SP 800-82 Rev 3 | Purdue zoning, network segmentation policy, declarative control mapping |
| CMMC L2 / NIST SP 800-171 | Policy engine gates every write; audit log with Ed25519 signatures |
| NIST AI RMF | Every decision mapped to Govern/Map/Measure/Manage; exportable |
| ISO 42001 | Governance lineage + human oversight + risk classification |
| ISO 14224 | Asset taxonomy used for metadata decoration |
├── src/industrial_agents/ # Python package
│ ├── agents/ # 10 industrial agents
│ ├── tools/ # Protocol adapters (OPC UA, MQTT, PI, Snowflake, …)
│ ├── orchestration/ # AutoGen group-chat + routing
│ ├── governance/ # Lineage, signing, OPA, PII redaction
│ ├── security/ # Purdue zoning, mTLS, secrets vault
│ ├── observability/ # OpenTelemetry, Prometheus
│ └── ui/ # Streamlit operator copilot
├── config/ # ISA-95, UNS taxonomy, Purdue zones, CMMC, LLM config
├── docker/ # docker-compose + Dockerfiles
├── policies/ # OPA rego policies
├── data/synthetic/ # Reproducible synthetic UNS dataset
├── examples/ # Five canonical Jupyter notebooks
├── benchmarks/ # Industrial Agent Benchmark (IABENCH-v1)
├── docs/ # MkDocs site source
└── legacy/ # Archived original SDLC demo
| Task | Metric | Claude Sonnet | GPT-4o | Llama 3.1 8B |
|---|---|---|---|---|
| Root-cause attribution | P/R F1 | TBD | TBD | TBD |
| Tacit-knowledge retrieval | nDCG@5 | TBD | TBD | TBD |
| Safety guardrail compliance | Block-rate | TBD | TBD | TBD |
| Hallucination rate | % | TBD | TBD | TBD |
Headline benchmark results will be published alongside the v0.1.0 release. See benchmarks/industrial_agent_benchmark.md for the full task specification and benchmarks/runner/ for the reproducible harness.
See CONTRIBUTING.md. This project follows the Contributor Covenant v2.1.
If you use this framework in research, please cite:
@software{misra2026industrialagents,
author = {Misra, Adris},
title = {Industrial Agentic Intelligence Framework},
year = {2026},
url = {https://github.com/adris-misra/multi-agentic-framework},
version = {0.1.0}
}MIT © 2026 Adris Misra