When Datta Kaligotla and Oscar Rangel set out to build an AI company, they picked one of the messiest problems in enterprise software: getting factory systems to talk to each other. Their startup, Faraday, emerged from Y Combinator's fall 2026 batch with a bold pitch — AI agents that don't just suggest actions but execute them across the patchwork of systems that run modern manufacturing.
The San Francisco company published a benchmark last month showing its agents outperformed standalone foundation models on multi-system industrial tasks by more than 20 percentage points. That timing wasn't accidental. SAP, Siemens, and Honeywell all rolled out competing agent products within days of each other in October, and according to BCG's 2026 report, 44% of companies now deploy AI somewhere in their supply chains.
But Faraday is betting on a different architecture. Where the incumbents embed agents inside their own software suites, Kaligotla and Rangel built what they call a "shared runtime" — infrastructure that ties together the fragmented landscape of enterprise resource planning systems, manufacturing execution software, product lifecycle management tools, warehouse systems, and even factory-floor robots. The agents learn through reinforcement environments trained on live operational data, tuning themselves to each company's approval workflows and exception rules.
It's an ambitious technical claim from a team of two.
What the Platform Does (and Claims to Do)
Faraday's system handles production scheduling, engineering review, quoting, inventory optimization, predictive maintenance, robot deployment, and three-way invoice matching for accounts payable. The company's use-case guide, published in October, walks through a typical workflow: agents ingest read-only data from an ERP system like Epicor Prophet 21, recompute inventory min-max levels weekly, generate buyer plans with the necessary approvals, then write results back to the source system.
The architecture is vertically integrated, according to the company's materials, meaning it doesn't just recommend actions but analyzes, executes, and automates workflows end-to-end. Customer-specific reinforcement learning environments let the agents adapt to constraints that vary wildly across manufacturers — one company's approval process might require three sign-offs, another might route through regional managers first.
"An agent earns its place only when the work survives contact with the real world," Kaligotla wrote in a September profile. "An agent earns its place in the operation one correct handoff at a time."
That caution reflects the founders' backgrounds. Kaligotla spent time at Microsoft and Boeing before becoming a founding engineer at Handshake AI, with earlier stints as an enterprise quant at Pretium Capital. Rangel built factory operations systems at Worthington Steel, shipped medical devices at Stryker, and worked on projects for NASA astronauts. They've seen enough enterprise software implementations to know that "integration" often means duct tape and workarounds.
The Benchmark Wars
Faraday released what it calls the Faraday Industrial Benchmark in October — a public evaluation suite with 74 task episodes across 33 scenario families spanning manufacturing, supply chain, logistics, engineering, and ERP finance workflows. Unlike general-purpose AI benchmarks, this one includes state-checking and authorization verification, testing whether agents can navigate the kind of multi-system exceptions that happen constantly in real operations: supplier delays trigger production rescheduling, which cascades into accounts payable reconciliation.
The company's verified results show its Cage agent, built on GPT-5.5, scored 96.2 out of 100. Anthropic's Opus 4.7 hit 75.2, standalone GPT-5.5 reached 71.7, and Google's Gemini 3.1 Pro landed at 61.4. Faraday was careful to note the benchmark runs in a synthetic sandbox and makes no claims about live-operations safety.
Benchmarks, of course, are only as meaningful as the tasks they measure. And the company's website and research data mention ongoing development but no specific deployment or customer verification yet.
The Incumbents Move Fast
Industrial software giants rarely move in lockstep, but October brought a flurry of agent announcements. SAP expanded its Joule platform with specialized agents for manufacturing, logistics issue resolution, and asset operations, targeting general availability in the fourth quarter. One customer, Doehler, cited one-click logistics decisions that recovered a third of the time previously spent on exception handling.
Honeywell Technologies launched Forge Cognition on October 14 with NVIDIA and Dell Technologies. Pilot results showed a 90% reduction in unnecessary alerts and up to 10% energy savings across multiple sites, according to the company. "The physical world demands more than a powerful AI model," said Suresh Venkatarayalu, Honeywell's chief technology officer. "It requires AI to understand how systems actually operate and interact."
Siemens had expanded its Industrial Copilot family earlier, introducing the Fuse EDA AI Agent in July. Microsoft updated its Foundry Agent Service documentation in September with managed runtimes and guardrails for production deployments. Schneider Electric unveiled "agentic manufacturing" capabilities with Azure AI at Hannover Messe in April.
The pattern was clear: every major player saw the same opening at roughly the same time.
A Crowded Seed Stage
Faraday isn't the only startup chasing this opportunity. Two other AI-agent companies targeting manufacturing raised seed rounds in October, according to Axios Pro reporting. Evos secured $9 million in pre-seed funding for AI operators serving small and mid-sized industrial companies. A startup called Forgepoint raised $9.8 million led by Primary Venture Partners and Andreessen Horowitz.
DeepFabric announced general availability of its AI agent platform for end-to-end supply chain in July, citing customers including NFI, Kenco, HelloFresh, TwinMed, Merchants Fleet, and Weber. Blue Yonder and other supply-chain software vendors added agent features throughout the year.
Even manufacturers themselves are building. BMW Group deployed a multi-agent system called Alconic in purchasing operations and is piloting humanoid robots at its Leipzig plant, which the automaker said in September would provide "a foundation in place for AI agents to handle demanding tasks autonomously." Foxconn announced AI factory blueprints with NVIDIA in March.
Faraday hasn't disclosed funding beyond Y Combinator's backing, and the team remains at two people, according to the accelerator's listing. YC partner David Lieb led the investment.
Adoption Outpaces Governance
The appetite for industrial AI is running ahead of the infrastructure to manage it. Deloitte reported in July that 40% of surveyed manufacturers now use generative or agentic AI, while an MHI industry report found 53% using AI with 88% expecting adoption within five years.
But a BCG study published in October found a troubling gap: 42% of companies expect agents to make autonomous decisions by 2030, yet only 5% reported sufficient governance controls in place.
Gartner warned in June that agentic AI elevates infrastructure and operations risks due to nondeterminism, dynamic tool invocation, and multi-agent complexity. ServiceNow introduced something called AI Passport in October to provide verifiable agent permissions at runtime. The European Union's AI Act phases in obligations for high-risk AI systems between December of next year and August 2028.
The skills gap compounds the urgency. U.S. manufacturing posted 522,000 job openings in August, according to the National Association of Manufacturers, with 1.9 million positions expected to remain unfilled over the next decade. Jensen Huang, NVIDIA's CEO, said in June that "AI factories are the infrastructure of the new industrial revolution," as reported by the Associated Press.
Gartner estimated in August that industrial AI represents a $300 billion opportunity as software and agent layers expand at the expense of traditional operational-technology hardware, disrupting incumbent vendor models in the process.
Still Proving It Out
Faraday posted on LinkedIn within the last week that it "spun up this workflow to take a customer RFQ through engineering review and quote preparation," with agents working "across your documents and existing systems, configured around your processes and approval rules."
The company hasn't disclosed production deployments or plans to expand beyond its founding team. That might sound lean for a startup taking on SAP and Siemens, but perhaps the founders see it differently. In industrial software, credibility comes from surviving contact with real factories, not from headcount.
Whether two people can build and scale a platform that orchestrates the messy reality of manufacturing remains an open question. But if the agents work as advertised, Faraday won't stay a two-person company for long.
