On Tuesday, March 3, 2026, Berlin-based Juna AI made its move. The company unveiled what it calls an Agentic Factory OS—essentially, an operating system for coordinating multiple AI agents across industrial production lines. The timing, as these things go, is no accident.
Juna has been building toward this since securing $7.5 million from Kleiner Perkins on November 18, 2024, a seed round that also pulled in Norrsken VC and a handful of notable angels, including John Doerr himself. The company landed on Forbes' Cloud 100 Rising Stars list last September, cited for its reinforcement learning work across "dozens of customers" spanning Europe, the U.S., and Asia. Now it's making a bigger claim: that the future of factory automation isn't about standalone tools or conversational chatbots, but about coordinated AI agents that share context, negotiate trade-offs, and optimize production in real time.
It's an ambitious bet—one that puts Juna squarely in the path of industrial giants already racing to define what AI-native manufacturing actually looks like. Siemens rolled out "Industrial AI Agents" last May and doubled down at CES in January with NVIDIA, unveiling an expansive vision involving nine copilots and something they're calling an Industrial AI Operating System. Microsoft launched its Factory Operations Agent in public preview around the same time last year. Salesforce is pitching Agentforce for Manufacturing. Even within the startup world, the "factory OS" language is catching on—First Resonance shipped ION Factory Operating System last June with over 60 deployments; Tetra Pak introduced its own Factory OS late last year.
Which raises the question: What makes Juna's version different? And can a small Berlin team with fewer than 50 employees carve out space in a market where established players have decades of customer relationships and massive R&D budgets?
Four Agents, One Shared Brain
Juna's platform centers on what it calls an Orchestration Layer—a connective system that integrates sensors, PLCs, SCADA, historians, MES, and ERP data into a unified semantic foundation. On top of that sit four pre-built agents, each handling a discrete operational domain.
The Production Scheduling Agent tackles dynamic planning. Juna claims it delivers 8% higher utilization, 5% improvement in on-time-in-full delivery, and 7% lower energy costs. The Process Engineering Agent automates monitoring and root-cause analysis; the company says it cuts time spent on routine tasks by 80% and reduces irrelevant alarms by 70%. The Control Agent offers real-time parameter recommendations aimed at yield and throughput—10% yield gains, 5% throughput lift, and 15% energy savings, according to Juna's figures. Finally, the Resource Intelligence Agent tracks material and energy flows, with claimed impacts of 18% lower energy costs and 9% fewer CO2 emissions.
These are big numbers. They're also unaudited and drawn from case studies Juna published itself, which means the usual caveats apply.
The agents don't operate in isolation. That's the architectural distinction Juna is pushing. A scheduling decision might ripple into energy management; a quality alert could trigger adjustments in process control. The platform is designed so agents can "see" each other's context and coordinate across objectives—something that requires more than bolting a chatbot onto a legacy system.
Human oversight remains central, at least in principle. Operators define objectives, constraints, and guardrails. Automation levels are configurable. There's a Factory Chat interface—what Juna describes as "production-aware intelligence," not a generic LLM wrapper—that lets users query live and historical production data in natural language. ("Why did yield drop on the night shift?" or "Compare energy use across last week's batches.")
Customers deploy agents through a self-service Agent Builder, which includes an SDK for orchestrating models and workflows. Background agents run continuously, monitoring KPIs, generating shift reports, flagging quality correlations. It's meant to feel like an operating layer, not a collection of point solutions.
The Proof Is in the Margin
Juna's go-to-market hinges on case studies with eye-catching financial claims. A dairy manufacturer using the Control Agent on a spray-dryer reportedly increased profits by €1.5 million annually, per a case study the company published last November. A carbon-black producer—Deutsche Gasrußwerke, which Juna names publicly on its homepage—deployed the Scheduling Agent and saw a €5.3 million annual contribution margin increase, according to a case study the company published. That work involved a live digital twin layered over ERP, real-time signals, logistics, and sustainability constraints. The optimization engine combines mixed-integer programming with AI, generating feasible schedules in under a minute.
CEO Peter Hartmann of Deutsche Gasrußwerke is quoted on Juna's site regarding scheduling impact, aligning with the €5.3 million figure. Juna also lists "typical results" of 6% throughput gains, 15% energy cost reduction, and 9% gross profit improvement on its homepage, though those numbers are undated and lack third-party verification.
The company is targeting process industries—chemicals, food and beverage, cement, pulp and paper, glass, pharma, and broader industrial manufacturing. These sectors share common pressures: razor-thin margins, aging expertise (retirements are accelerating knowledge loss faster than companies can document tribal know-how), and mounting sustainability mandates that increasingly link environmental outcomes to operational performance.
It's a sensible focus. Process manufacturing tends to be continuous or batch-oriented, data-rich, and already instrumented—meaning there's infrastructure to plug into. The question is whether Juna can move fast enough, and prove value consistently enough, to build a defensible customer base before incumbents catch up or larger platforms absorb the opportunity.
A Very Crowded Window
Juna is hardly operating in a vacuum. The agentic AI wave in manufacturing is real, and it's moving quickly.
Siemens announced its Industrial AI Agents initiative at Automate Detroit last May, promising orchestrated autonomous agents across the industrial value chain. By January's CES, the vision had expanded considerably—partnering with NVIDIA to introduce a sprawling ecosystem that included nine copilots and a Digital Twin Composer. Microsoft's Factory Operations Agent, launched in public preview last March, offers agentic assistants for frontline manufacturing through Azure AI Foundry and Copilot Studio. Salesforce is positioning Agentforce for Manufacturing as "digital labor" for commercial workflows and field operations.
Even within the narrower "factory OS" category, the terminology is converging fast. First Resonance released ION Factory Operating System with agentic "ION Intelligence" last June, claiming more than 60 company deployments. Tetra Pak launched its own Factory OS around the end of last year, positioning it as an AI-ready platform for food and beverage with standardized data collection. UK-based Matta and Aris Machina are also staking claims on "manufacturing OS" and "agentic OS" frameworks.
Incumbent players bring different strengths. AspenTech's Plant Scheduler has long been a standard for continuous and batch detailed scheduling in process industries, though it predates the agentic era. Yokogawa achieved what it describes as the world's first official adoption of autonomous reinforcement learning control at an ENEOS Materials plant in 2023 and has been expanding deployments with partners like Aramco. Juna's Control and Process Engineering agents compete directly with this category of advanced process control and RL systems.
It's worth noting that "agentic" has become something of a buzzword—every vendor wants to claim they're building agents, not just software. The distinctions matter, though. True agentic systems involve goal-driven behavior, contextual reasoning, and the ability to coordinate across multi-step decisions. Not all platforms claiming the label deliver that.
German Hosting, ISO Claims, and Data Sovereignty
Juna emphasizes data sovereignty and security—a selling point in regulated industries and European markets wary of cloud hyperscalers with U.S. jurisdiction. The platform is hosted in Germany, with isolated customer environments running on separate servers. The company claims ISO 27001 certification and GDPR compliance, though it doesn't provide a registry link or certificate scope on its public pages. (A minor omission, perhaps, but one that might raise eyebrows among procurement teams used to verifying such claims.)
Integration depth is critical in process manufacturing, where decades of legacy systems coexist in uneasy alliances. Juna's platform includes native connectors across OT and IT stacks, aiming to eliminate the stitching work that typically slows AI adoption. The go-to-market is anchored in "Proof of Value" engagements designed to show measurable impact within weeks, then scale across production lines and plants.
The roadmap focuses on expanding pre-built agents, deepening integrations, and broadening access to the Agent Builder for in-house extensibility. Longer-term, Juna envisions cross-floor agent coordination in real time—agents that negotiate resource allocation, resolve conflicting objectives, and optimize holistically across facilities. Whether that vision materializes depends on execution, customer trust, and whether the platform can handle the messy complexity of real-world production environments.
A Founders' Second Act
Juna was founded by Matthias Auf der Mauer and Christian von Hardenberg, both with prior exits under their belts. Auf der Mauer previously founded and led AiSight, which Sensirion acquired in 2021. Von Hardenberg served as CTO at Delivery Hero and worked at Rocket Internet before that. The company was officially registered in Germany on September 26, 2024, making it a relatively new entrant despite the founders' pedigrees.
Employee count remains small. LinkedIn lists the company at 11 to 50 employees as of mid-February; third-party directory Dealroom shows 2 to 10, highlighting the opacity that early-stage companies often carry. It's not unusual—headcount can fluctuate, and different platforms use different methodologies—but it underscores just how lean the operation is relative to its ambitions.
No additional funding beyond the November seed round has been announced. That $7.5 million will need to stretch while Juna proves it can convert pilots into long-term contracts and expand beyond its initial customer base. The Forbes recognition last September noted "dozens of customers," though the company hasn't disclosed names beyond Deutsche Gasrußwerke.
The Bet on Orchestration
Juna's launch comes as broader momentum builds around agentic AI in manufacturing. NVIDIA published an "Agentic AI in the Factory" section in its AI Factory Design Guide this February, outlining operational patterns and blueprints for agentic deployments. Trade publications and vendor blogs have forecast 2026 as a breakout year for agentic adoption, with predictions of self-healing supply chains and tighter integration between shop floor and design studio.
Whether those predictions hold depends on whether platforms like Juna's can deliver on the promise of orchestrated intelligence—and whether enterprises trust multi-agent systems enough to hand them the keys to production. Process manufacturers are notoriously risk-averse. Downtime is expensive. Quality failures can trigger recalls. Energy inefficiencies hit margins directly. The stakes are high, which means the bar for proof is high too.
Juna is betting that the old model—isolated optimization tools that don't talk to each other—is finally running out of runway. That coordination, not just automation, is what unlocks the next level of operational performance. It's a compelling thesis. The challenge, as always, is turning thesis into reality at scale.
For now, the company has a product, a handful of case studies, and a market window that's filling up fast. Whether that's enough remains to be seen.
