How the New AI Agents for Manufacturing From Tulip, Siemens and Honeywell Compare

Three vendors announced AI agents for manufacturing within two weeks this October, and none of them published a price. Tulip expanded its Composable AI Agents on October 6. Honeywell Technologies introduced Forge Cognition on October 14. Siemens and NVIDIA showed agent-based production planning at Siemens' Erlangen electronics plant on October 19. The three platforms go after different problems. Tulip is aimed at frontline apps and the workflows around them, Honeywell at facility and building operations, and Siemens at production planning and ramp-up.

How we built this comparison: it draws on the vendors' primary releases and a small number of press reports, all linked below. We did not test any of the products. Every performance figure in this article is the vendor's own. The comparison is independent of sponsorship.

What are the new AI agent platforms for manufacturing?

In brief, Tulip offers composable agents inside its frontline operations platform, Honeywell offers a self-contained agentic AI and simulation platform for facilities, and Siemens offers planning agents that are checked against a digital twin of the whole plant. All three tie AI agents to operational data and keep a human in the approval loop, but the jobs they do differ.

Tulip: Composable AI Agents and MCP connectors

Tulip made its announcements at Operations Calling 2026 in Boston and framed them as capabilities "that put AI to work for the people who run the factory floor," according to the company's press release. The release says Composable AI Agents were "Introduced at Operations Calling 2025" and "can be used in even more floor operations, always with a preserved operational record for human-in-the-loop quality assurance." The October release gives the agents no explicit availability label. They did, however, appear in Tulip's LTS16 release on July 1, 2026 marked as Beta, described as agents "trained on your own SOPs, validation frameworks, and data models to accelerate every step from solution design to production."

Tulip announced several pieces around the agents:

• MCP Connectors (in preview mode): native Model Context Protocol connectors that work in both directions. According to the release, "external AI tools and coding agents can read from and act on Tulip apps, tables, and workflows, and Tulip agents can call external systems the same way."

• Industrial Connectivity (preview): an operational data layer that brings machine data from existing drivers, historians and a unified namespace (UNS) into Tulip and joins it with human context.

• Factory Playback: first introduced with NVIDIA last year, it "turns local edge server camera streams and advanced vision-language models into a complete, replayable history of the floor." Tulip says new human activity detection shows where time goes, whether a procedure was followed and whether a part was assembled to spec.

• Enterprise Library: a single library of apps and assets on a customer's Center of Excellence instance, due to ship later this year.

CEO Natan Linder set out the company's position in the release: "AI has one job, to unlock human potential. Humans are not going away; they will keep working differently and better."

Honeywell Technologies: Forge Cognition

Honeywell's October 14 release describes Forge Cognition as "a self-contained agentic AI and simulation platform for facility operations." It combines real-time digital twins with AI agents across buildings, campuses and industrial operations. Most of the software stack comes from NVIDIA: Nemotron open models, Omniverse libraries for simulation, Cosmos open models with the Metropolis VSS Blueprint for vision AI, and RTX PRO GPUs. Dell supplies the infrastructure.

Suresh Venkatarayalu, Honeywell Technologies' chief technology officer, said in the release that the platform "combines AI with operational context and engineering expertise, opening a new era of intelligent operations where AI can analyze, recommend and act alongside people."

Siemens and NVIDIA: Intelligence Center X agents at Erlangen

The Siemens release, "Siemens partners with NVIDIA to advance AI-driven, autonomous factories," follows the companies' January 2026 CES announcement that they would build "the world's first fully AI-powered manufacturing sites." At Erlangen they demonstrated two things: virtual production ramp-up and agent-based production planning.

The NVIDIA blog post on the event says "Siemens Intelligence Center X agents, built with the NVIDIA Factory Operations Blueprint, use factory data and planning tools to recommend options." The agents compare materials, line capacity, delivery commitments, and changeover time and cost. Siemens' release says that when new orders arrive, the system "automatically generates various production scenarios and shows planners the impact of each option based on defined KPIs," and that "Before implementation, these proposals are validated using a Digital Twin of the entire production facility."

On inspection, NVIDIA Cosmos world models produce the training data. Per Siemens: "AI now generates photorealistic representations of PCB assemblies from their digital twins to train the vision systems, enabling engineers to prepare inspection before physical prototypes are available."

Intelligence Center X itself is not new this month. Digital Engineering 24/7 reported that Siemens announced it at Realize LIVE Americas 2026 in Detroit. It combines Mendix low-code, Graph Studio and AI Studio, which comes from the Rapidminer portfolio. That report gave no general availability date and no pricing.

The plant matters for context. Siemens says the Erlangen site was established in 1971 and employs approximately 1,000 people, and that "more than 100 AI algorithms are already in active use across production, logistics, quality inspection and maintenance." Cedrik Neike, who heads Siemens Digital Industries, cites "an annual productivity increase of seven percent." The site is set for further expansion through 2030, with Siemens investing €500 million in the transformation. NVIDIA's blog describes that same €500 million as investment in the Erlangen Technology Campus. This is a mature, heavily instrumented showcase plant, not a typical mid-market shop.

Where do factory AI agents run: cloud, on-prem or edge?

Honeywell gives the clearest answer. Forge Cognition runs on-premises, and the release states: "Dell Technologies provides the AI infrastructure underpinning Forge Cognition through the Dell AI Factory with NVIDIA, including Dell PowerEdge XE servers, Dell AI Data Platform data and storage engines, Dell cyber resiliency solutions and support from Dell Services." It adds that "The platform uses NVIDIA RTX PRO GPUs and is expanding to accommodate far-edge deployments with NVIDIA DGX Spark."

Tulip's most explicit on-site element is Factory Playback, which works from "local edge server camera streams." The October release does not give a deployment model for the agents or the MCP connectors.

Siemens' release does not say whether the Erlangen agents run in the cloud, at the edge or on-premises.

For a plant with data residency requirements or an air-gapped network, this is the first thing to settle. If a vendor runs inference on dedicated GPU servers in your building, you will need rack space, power, cooling and IT support budgeted alongside the license. If inference runs off-site, you will need a clear statement of which production data leaves the plant.

What systems do the agents connect to?

Tulip has the most detailed integration story. Its MCP connectors run both ways, so outside AI tools can act on Tulip apps and Tulip agents can call outside systems. Industrial Connectivity pulls machine data from existing drivers, historians and a UNS. For plants already running Tulip apps on the line, this is the most direct route. Bidirectional "act on" access also means permissions and audit scope need to be designed before any connection goes live.

Honeywell says Forge Cognition coordinates actions "across supported existing systems." The release does not list which systems are supported, so buyers should ask for that list.

Siemens' agents apply approved changes "through the connected planning system," according to NVIDIA. The demonstration took place in a Siemens plant on Siemens tooling. How well the approach carries over to a plant running a different ERP, MES or scheduling stack is not addressed.

Which factory AI agent platforms are available now?

None of the three is described as generally available across the board. Each one is at a different stage.

• Tulip: Composable AI Agents shipped as Beta in LTS16 in July. MCP Connectors and Industrial Connectivity are in preview. Enterprise Library ships later this year.

• Honeywell: Forge Cognition is in pilot at six named sites: NVIDIA's Voyager campus, the Fairmont Hotel in Dubai, NEX Mall in Singapore, Vanderbilt University, and Honeywell's own campuses in Bangalore and Richardson, Texas. Per the release, "Customers can engage Honeywell Technologies today to identify and scope Forge Cognition deployments, with capabilities expanding over time." Scoping is not the same as a shipping SKU.

• Siemens: the agents are running in production testing at Erlangen. The release says nothing about commercial availability for customers.

Five of the six Honeywell pilot sites are hotels, malls, universities and corporate campuses. The release positions the platform for industrial operations too, but none of the named pilots is a production line.

Can AI agents in a factory act without human approval?

All three vendors keep a person in the loop for now, and they do it in different ways.

• Honeywell: "Customers can operate Forge Cognition in manual, semi-autonomous or autonomous modes. Operators set the level of oversight and control and can override it at any time." The autonomous mode is listed, but Venkatarayalu told Fortune that "We have not opened up an autonomy right now" and that "For me, I wouldn't deploy these agents without my control system in place."

• Siemens: per NVIDIA, "The production controller reviews the trade-offs and approves the plan before the agent applies permitted changes through the connected planning system." The agent recommends options, a person approves, and the agent then makes only changes it is permitted to make.

• Tulip: agents run "always with a preserved operational record for human-in-the-loop quality assurance."

On regulated lines (FDA, AS9100, customer-audited automotive), the preserved record matters as much as the approval step. Ask each vendor for a sample audit trail: what the agent saw, what it recommended, who approved it, and what changed in which system.

How should operators read the vendor numbers?

Every figure below comes from the vendor and is a projection or a pilot result. None comes from independent measurement.

Siemens says that for new printed circuit board assemblies, which are the core of the drives and frequency converters made at Erlangen, "virtual production preparation can reduce engineering effort by up to 85 percent and cut time-to-market by up to 40 percent." It also says "end-to-end agent-based production planning can enable cost savings of up to 20 percent." All three figures are "up to" values for one product family at one plant that already runs more than 100 AI algorithms. Treat them as an upper bound from a best case, not a planning assumption.

Honeywell's pilot figures come from NVIDIA's Voyager campus: "Up to 90% reduction, on average, in unnecessary alerts," an identified "up to 10% potential reduction in energy and utility consumption," and a comfort score held above 95% (a weighted average of temperature, humidity and CO2). These are building metrics. The alert-fatigue result may carry over to plant utilities and maintenance alarming, but it is not evidence of production-line throughput or quality gains. The energy figure is a potential reduction that was identified, not one that was realized.

Tulip published no performance figures.

If you are building a business case on alert reduction or uptime, model the value from your own numbers. Our downtime cost calculator separates lost contribution margin from incremental recovery costs, which keeps an agent's projected benefit from being double-counted.

What have the vendors not disclosed?

What should a plant manager ask each vendor?

• Where does inference run, and where does our data go? Get it in writing for each component, not just the platform as a whole.

• What hardware do we need? For on-prem options, ask for server count, GPU class, power and cooling for a plant of your size, and who supports it.

• What does it integrate with today? Ask for the supported list for your MES, historian, ERP and scheduling system, and whether each integration is shipping, in preview or on a roadmap.

• What can the agent do without a person? Find out which actions need approval, who approves them, and whether the permission model can be set per line or per system.

• What does the audit trail look like? Ask for a sample record you could hand to an FDA, AS9100 or customer auditor.

• What is the commercial model? Licensing basis, hardware ownership, services, and what a pilot costs compared with a production rollout.

• What has been measured on a production line like ours? Separate vendor projections and building pilots from results measured on comparable lines.

Bottom line for the capex budget

These are three different bets, not three versions of one product. Tulip fits plants already building frontline apps that want agents and MCP access on top of them. Honeywell fits facility-wide operations, meaning utilities, energy and alarm load, with a defined on-prem hardware stack. Siemens is showing what planning agents can do in a highly digitized plant on its own tooling, with no stated path to buy it yet. Until pricing and general availability dates are published, the sensible step for most mid-market plants is a scoped pilot with clear exit criteria, not a platform commitment. If you are weighing agents against a core MES decision, start with our MES buying guides and make the agent layer prove its value on top.

Related reading

• When Should a Mid-Size Plant Register Its Own Apprenticeship Program?

• Is Your Motor Supplier Ready for the January DFARS Magnet Rule?

• Can 3D Printed Solid Rocket Motors Reach Full Rate in 2027?

• How to Choose Predictive vs Preventive Maintenance for Each Asset

• Why Your OEE Number Hides Most of Your Lost Capacity

• More Industrial IoT coverage

Sources

• Tulip Brings AI to the Factory Floor at Operations Calling 2026 (Tulip press release, Oct 6, 2026)

• LTS16 Is Here! AI Agents, Connector Versioning, more (Tulip Community, July 1, 2026)

• Honeywell Technologies Launches Forge Cognition with NVIDIA and Dell Technologies (Honeywell press release, Oct 14, 2026)

• Exclusive: Honeywell Technologies launches a 'brain behind the building' with Nvidia and Dell (Fortune, Oct 14, 2026)

• Siemens partners with NVIDIA to advance AI-driven, autonomous factories (Siemens press release, Oct 19, 2026)

• Siemens Factory Erlangen: Where Industrial AI Becomes Reality (Siemens press feature)

• NVIDIA Lands in Europe for GTC Berlin (NVIDIA blog, Oct 19, 2026)

• Siemens Launches Industrial-Scale, Agentic AI Orchestration Platform (Digital Engineering 24/7, June 2, 2026)