I’ve been spending a lot of time thinking about what it really takes for AI to participate in warehouse operations—not just answer questions about them.
That work is open in MAIW v3.0.0 — Multi-Agent Intelligent Warehouse:
github.com/NVIDIA-AI-Bluepri…
MAIW is a dev example for governed agentic warehouse operations.
The core idea is simple: AI can be adaptive in how it reasons, while operational authority remains explicit, deterministic, and auditable.
The blueprint brings together:
Nemotron models, SOP-driven agents, Warehouse State, MCP interoperability, deterministic governance, human approval, guarded execution, and closed-loop outcome verification.
The lifecycle is:
Observe → Reason → Recommend → Govern → Approve → Execute → Verify
What matters to me is where this can go next.
I’m now applying the same architecture to a focused yard-management use case, combining perception, OpenUSD simulation, operational context, and NVIDIA Cosmos to evaluate possible trailer and dock decisions before acting.
The bigger opportunity is to give developers a reusable foundation for connecting AI reasoning + physical operations + governance + simulation + learning.
Love any feedback and a collaborative engagement around those ideas.
1
16,732
Berlin, let’s put open models to work!
I’ll be at the @nebiusai booth at @nvidia GTC Berlin, demonstrating how we turn open models into specialized warehouse agents through the Intelligent Warehouse Fine-Tuning Flywheel.
See how warehouse experience becomes training data, how fine-tuning builds domain expertise, and how evaluation and guardrails guide continuous improvement—from inventory and labor planning to responding to operational disruptions.
Powered by NVIDIA AI technology and Nebius infrastructure, the demo connects model customization with a practical goal: agents that understand warehouse workflows, use tools effectively, and operate within business constraints.
Building agents for retail, CPG, manufacturing, or supply chains? Come see the flywheel in action and let’s compare ideas.
Meet me at the Nebius booth. Let’s build the intelligent warehouse of the future!
2
4
6,048
Just came off stage at ALL IN 2026 in Montréal.
Great panel with Mila, IVADO Labs, Kinaxis & Jamieson Wellness on AI in logistics — speed, resilience, cost.
Autonomous operations aren’t coming. They’re here. The question is who builds the stack first.
Proud to represent NVIDIA. Montréal delivered.
1
4,340
Looking forward to this conversation at Stanford OpenLab with @karenhsumar , BASES and the Stanford OpenLab community.
The next generation of physical AI will not be defined simply by putting more intelligence into robots or connecting LLMs to machines. The real opportunity is to build closed-loop decision systems that can understand the current state of the physical world, predict how that world may evolve, and decide what should happen next within explicit safety, policy, and operational boundaries.
In the work I have been developing around intelligent warehouses, this distinction has become central: the digital twin helps us understand what is happening now, the world model helps us reason about physically plausible futures, and the decision layer determines which future we should pursue. Agents can reason, plan, and propose—but reasoning is not authority. Trustworthy autonomy comes from combining agentic AI, physical world models, optimization, and governed execution.
Excited to discuss what this means for builders creating the next generation of agents, robots, digital twins, and autonomous systems—and why the path from AI to autonomy is ultimately a problem of decision architecture, not just model intelligence.
luma.com/5k4twswi
2
4,810
If you’re heading to #ALLIN2026 next week, I invite you to stop by the Agora stage on September 17th to catch me in a panel session, AI in Logistics and Supply Chains: Speed, Resilience, and Cost.
Looking forward to diving into how we can leverage these innovations in AI to build smarter, more resilient, and cost-effective supply chains.
I'll be joining : Manik Sharma, Emma Frejinger Andre Teixeira and Enrico D. Cremonese for what should be a great conversation on building supply chains that are faster, more resilient, and smarter.
I look forward to seeing you next Thursday, September 17, from the Agora Stage at 1:15 p.m. ET!
ALL IN | 2026 => nvidia.com/en-us/events/all-…
81
I’ve been building an Operational Intelligence Lab — a living developer environment for exploring how AI services can understand dynamic physical systems, simulate possible futures, orchestrate trusted models, act safely, and continuously learn from operational experience.
The warehouse is the first proving ground.
The work brings together:
Dynamic Agentic Graphs → DES + Monte Carlo → World Models → Agents + Optimization → Operational Experience → SFT/RL → Continuous Learning
This isn’t about proposing another WMS or one monolithic architecture. The goal is to build and eventually open-source a developer blueprint: tools, patterns, experiments and reference implementations for the next generation of operational AI.
I’m documenting the journey openly through a series of articles on Operational Intelligence Lab.
Here is my latest notes: open.substack.com/pub/tarikh…
More experiments, architecture, code—and eventually the open-source blueprint—coming.
The next frontier of AI isn’t only answering questions. It’s understanding the world, exploring what could happen, making better decisions, and learning from what actually happened.
3
10,865
Assembling operational context from a relational database takes 4.2 seconds. From a native graph structure, it takes 50 milliseconds.
That is not a performance optimization. It is the difference between a system that can and cannot operate at the speed warehouse decisions are actually made.
The Agentic Graph — G=(E,R,T,L,D) — is the representation layer that makes real-time operational AI possible. Five components: entities carrying live state, active relationships between them, real-time telemetry attached to both, a four-hour event trace, and a history of decisions made in comparable situations along with their outcomes. Not a knowledge graph that was built once and periodically refreshed. A continuously updated model of what is operationally true right now.
What the agent reasons over is not the full graph — it is a dynamically assembled subgraph called the Operational Context Graph, extracted at 100:1 compression by the Operational Context Builder in under 50ms. The OCB identifies the entities causally connected to the current disruption, traverses their dependency chain, enriches the subgraph with telemetry and event history, and delivers a decision-ready context. The agent never sees the whole facility — only the part that matters for this decision.
One architectural detail worth adding: the graph is hybrid. Entities, relationships, and telemetry are structured graph data — precise, queryable, inspectable. The event log and decision history are embedding-indexed — because finding comparable past situations is a semantic similarity problem, not a structural query. The graph handles now. The vector index handles memory.
An AI agent with the right context makes better decisions. That is the entire argument. Full post here: lnkd.in/g5jRW76E
#SupplyChain #AIEngineering #GraphML #WarehouseAI #EnterpriseAI #OperationsResearch
🤖 Made with AI
2
2
116
One of the biggest bottlenecks in advancing Agentic AI isn’t the reinforcement learning algorithm itself.
It’s the engineering complexity.
Every time researchers introduce a new reward function, estimator, rollout strategy, or training pipeline, they often end up modifying thousands of lines of distributed infrastructure instead of focusing on the science.
This is why I found NVIDIA’s new Molt framework particularly interesting.
Instead of treating RL infrastructure as a massive software engineering problem, Molt takes a different approach:
1- Pure PyTorch-native architecture
2- Compact enough to understand end-to-end
3- Asynchronous training for multimodal and MoE agents
4- Strict policy consistency (never trains on tokens it didn’t generate)
5- Comparable performance to significantly more complex distributed training stacks
To me, this represents something bigger than another RL framework.
It reflects a broader shift in AI:
The future belongs to frameworks that optimize researcher iteration speed, not just GPU throughput.
This resonates strongly with the work I’ve been building around continuous learning flywheels for operational systems.
Whether we’re training warehouse agents, manufacturing copilots, or enterprise decision systems, success won’t come from a one-time fine-tuning pipeline.
It will come from rapid experimentation:
• Generate synthetic experiences
• Fine-tune efficiently
• Evaluate continuously
• Reinforce successful behaviors
• Deploy updated policies
• Repeat…
The faster we can complete this loop, the faster AI systems evolve.
The infrastructure itself becomes a competitive advantage.
Excited to see research moving in this direction. Simplicity is often the catalyst for innovation.
Paper from NVIDIA Research: Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning (arXiv): arxiv.org/pdf/2607.21653
#AI #AgenticAI #ReinforcementLearning #PyTorch #LLM #NVIDIA #OperationalIntelligence #MachineLearning #OpenSource #ContinuousLearning
2
104
Supply chain is the only enterprise AI domain that simultaneously demands high decision frequency, high consequence per error, high operational variability, and high interdependency. Each of those properties individually makes AI harder to build and deploy. Together, they constitute an operating environment that breaks the assumptions underlying most enterprise AI architectures. #logistics #operationalintelligence #Agenticsystems open.substack.com/pub/thdev/… – at Kalaoa, HI
86
Honored to participate in the annual innovation event hosted by @Prologis in the beautiful town of Charleston.
What stood out most was the energy in the room — leaders, operators, technologists, and innovators deeply focused on the future of industrial infrastructure, logistics, energy, and supply chain transformation. The conversations were not just about warehouses or facilities; they were about building the intelligent foundation powering the modern economy.
One of the most powerful realizations from the event is this: Real estate is no longer just physical space. When connected with AI, data, automation, robotics, and energy intelligence, it becomes a strategic platform that maps value across the entire supply chain ecosystem — from manufacturing and fulfillment to transportation, inventory flow, and ultimately customer experience.
The convergence of industrial real estate, digital infrastructure, and AI-driven operations will define the next era of global competitiveness. The opportunity ahead is massive.
A big thank you to the entire Prologis team for the outstanding hospitality, the world-class organization, and for bringing together such an exceptional group of leaders and innovators. Truly inspiring. @ToriDeems @wbodonnell @WalterKemmsies @richardteachout @avihou @FreightAlley #innovation #supplychain #AI #automation #physicalai #prologis #nvidia
Just back from an inspiring NVIDIA GTC 2026 week in San Jose—tons of great things happening around AI factories, agentic systems, and turning data into real-time decisions.
Now chilling in Napa, diving into The Decision Factory: A Novel about Decisions Under Uncertainty by Adam DeJans Jr. and John Brandon Elam (foreword by Warren B. Powell).
Love the business-novel style—like The Goal, it uses a gripping story (Fulcrum Logistics team facing uncertainty, cascading failures, and adaptive fixes) instead of dry equations. You watch characters wrestle with the gap between forecasts and reality, building “decision factories” that actually work when plans go sideways.
It makes sequential decision analytics and handling uncertainty feel practical and intuitive—no heavy math upfront, just relatable struggles and breakthroughs.
Post-GTC, it hits perfectly: the future of ops isn’t just bigger models—it’s scalable systems that thrive under uncertainty.
Highly recommend for anyone in AI deployment, operations, or data-driven leadership. Quick, engaging read that bridges theory and real-world execution.
Who’s read it? Or got other fable-style books that made tough concepts click?
#DecisionIntelligence #AIDecisions #GTC2026 #OperationsResearch #UncertaintyManagement #TheDecisionFactory
52
I’ve been running OpenClaw 🦞as my always-on personal AI assistant for a week.
It writes its own skills. It spawns subagents. It keeps running after I close my laptop.
And honestly? It made me a little nervous.
Not because it wasn’t productive — it was extremely productive. But because the safety model was entirely trust-based. Guardrails living inside the same process they were supposed to be guarding.
NVIDIA just solved that problem with OpenShell.
Here’s what clicked for me reading this blog: developer.nvidia.com/blog/ru…
The fundamental insight isn’t about agents being more capable. It’s about the runtime layer finally catching up to the agents.
OpenShell sits outside the agent — between the agent and your infrastructure. Out-of-process policy enforcement. The agent literally cannot override it, even if compromised. It’s the browser tab model applied to autonomous agents: isolated sessions, permissions verified before any action executes.
Three components that make this real:
1- The Sandbox — not generic container isolation. Built specifically for self-evolving agents that write their own code mid-task. Agents can break the environment without touching the host.
2- The Policy Engine — evaluates every action at the binary, destination, method, and path level. If an agent hits a constraint, it reasons about the roadblock and proposes a policy update. You approve. It evolves within boundaries you define.
3- The Privacy Router — keeps sensitive context on-device with local models, routes to frontier models only when policy allows. Decisions made by your policy, not the agent’s judgment.
And to deploy it with OpenClaw?
- openshell sandbox create --remote spark --from openclaw
- Zero code changes.
One thing interesting here about my colleagues, the team behind this — acquired from Gretel in 2025 — built their careers at NSA, AWS Macie (petabyte-scale data protection), and enterprise synthetic data infrastructure. This isn’t academic. It’s production-grade security thinking applied to the agent layer.
The infrastructure decisions made in the next 6–12 months will define what enterprise agent deployment looks like for years.
We’re not in the “AI assistant” era anymore. We’re in the agent runtime era.
And it’s just getting started.
#AgenticAI #NVIDIA #OpenShell #OpenClaw #NemoClaw #AIAgents #DeveloperTools #GTC2026 #EnterpriseAI #AIInfrastructure
2
93
What if an LLM could design your entire image generation pipeline — and optimize it with reinforcement learning?
That’s exactly what FlowRL (ComfyGen-RL) does. A new paper from NVIDIA Research and Technion introduces the first RL-based framework for automatically generating ComfyUI workflows that are tailored to your prompt and aligned with human preferences.
Here’s the problem it solves: Text-to-image has moved far beyond single models. The best results now come from complex multi-component pipelines — fine-tuned generators, LoRAs, upscalers, editing steps — all stitched together in tools like ComfyUI. But designing these workflows requires deep expertise and hours of manual experimentation.
Previous approaches like ComfyGen tried to automate this using LLMs, but they had a key limitation: they essentially memorized and replicated existing workflows from training data. 94% of their outputs already existed in the training set.
FlowRL takes a fundamentally different approach:
Stage 1 — Supervised fine-tuning on 500K prompt-flow pairs teaches the LLM the structure and vocabulary of workflows. A custom tokenization scheme reduces flow representations from ~1,500 tokens to ~200, enabling a 16x batch size increase.
Stage 2 — GRPO-based reinforcement learning drives the model toward higher-quality workflow regions. Instead of generating actual images during training (which takes ~1 minute per step), they train an ensemble of surrogate reward models on ModernBERT that predict image quality directly from prompt-workflow pairs.
To prevent reward hacking, they use ensemble variance as an uncertainty signal — any prediction the models disagree on gets zero reward.
The result? FlowRL generates genuinely novel workflows (0% overlap with training data vs. ComfyGen’s 94%), achieves a 60% win rate over ComfyGen on human preference metrics, and matches state-of-the-art prompt adherence on GenEval.
They also introduce a clever inference trick inspired by classifier-free guidance: extrapolating between the SFT and GRPO-tuned model logits at generation time to further boost quality.
This feels like an important direction — not just for image generation, but for the broader idea of using RL to optimize compound AI systems. As pipelines get more complex, we need automated ways to navigate the design space. FlowRL shows that’s not just possible — it outperforms human-curated workflows.
Paper: arxiv.org/pdf/2505.21478
#AI #GenerativeAI #TextToImage #ReinforcementLearning #GRPO #ComfyUI #NVIDIA #MachineLearning #DiffusionModels #Research
1
3
99
Replying to @nvidia
@NVIDIA is now the driving force behind open AI model innovation powering the ecosystem with Nemotron, Cosmos, Gr00t, BioNeMo, and Canary.
Nemotron → agent intelligence
BioNeMo → biopharma discovery
Cosmos → physical reasoning
Gr00t → robotics learning
Canary → speech & voice
AI’s next wave is running on NVIDIA. #OpenSource #nvidia #agenticAI #AI #Robotics
56
















