The future of software may be written by machines, but it will still fail for ordinary reasons. An API times out. A payment needs approval. A process waits overnight, wakes up, and discovers that one of its dependencies has wandered off. Intelligence gets the keynote. Memory gets the pager.
Jeu George has made a career in that less theatrical territory. As an engineer at Netflix, he helped create Conductor, a system for coordinating work across distributed services without forcing every team to invent its own machinery for state, retries and recovery. Today, as co-founder and CEO of Orkes, he is returning to the same problem with a new cast. AI agents can reason, branch, call tools and collaborate. They can also stall, repeat themselves and lose the thread. George's bet is that the age of agents will need an execution layer as much as the cloud era needed infrastructure.
This is a story about orchestration, which sounds like the sort of word one might hide behind at a conference buffet. In George's telling, it is agreeably concrete. A useful application is a chain of actions. Each action may succeed, fail, pause or require a person. Somebody, or something, has to remember where the chain is and what should happen next. The developer should not have to rebuild that memory every Tuesday.
The mechanic discovers code
George grew up in southern India and studied mechanical engineering. The route into software arrived during an internship at the Indian Space Research Organisation. It was there, he has said, that he first encountered programming inside an enterprise. The machines were different from the ones in his degree, but the attraction was immediate. He came to the United States for a master's degree, then joined Microsoft.
His Microsoft years crossed two imposing pieces of software: Windows and Office. This was the pre-cloud period, when a serious application often lived on a serious box and everyone behaved as though the box might never develop a personality. George later moved to the Bay Area and joined Netflix in late 2011. DVDs were still part of the scenery. The company was moving into streaming, cloud computing and original content, a change that required more than a handsome red interface.
At Netflix, thousands of services had to combine into experiences that appeared simple to a viewer. Preparing and delivering original programming added long-running media workflows with plenty of opportunities for delay and failure. Conductor emerged from that pressure. George worked with Viren Baraiya and Boney Sekh on a platform that made workflow definitions visible, durable and independent of any single programming language or cloud.
The company also gave him an operating model. George spent roughly seven or eight years there and often points to Netflix's culture of freedom and responsibility as an education alongside the technical work. Small teams carried unusual ownership. People were expected to make decisions, understand the result and move without ceremonial layers of permission. From the outside, that arrangement can look like a slogan printed on an expensive wall. George describes it as something he lived daily, a practice that sharpened both his engineering and his judgment.
Orchestration grew in precisely that environment. A central platform did not need to centralize every decision. It could give independent teams a common language for composing services while leaving them free to choose the right implementation. The paradox is attractive: a little shared structure creates room for more autonomy. When the platform remembers the sequence, records the outcome and handles the predictable failures, engineers can spend their attention on the part only their team knows how to build.
The important design choice was not merely coordination. It was confession. A workflow should admit where it is, what it tried, and why it failed. Distributed systems are otherwise accomplished liars. Each service can report that it performed perfectly while the customer continues staring at a spinner.
A workflow with somewhere to put the plot
Two startups, two kinds of traffic
The infrastructure résumé has an unexpected subplot. George also co-founded Pinkvilla with Nandini Shenoy, his wife, who wanted better access to Bollywood news while living in the United States. Pinkvilla grew into a substantial entertainment publication. George's public biography places him across both the consumer web and enterprise plumbing, a pairing that is more coherent than it first appears. Publishing teaches brutal lessons about traffic, discovery and what users actually click. Infrastructure teaches equally brutal lessons about what happens after they click it.
After Netflix, George worked as an architect at Uber, including on Uber Marketplace. Then he reunited with colleagues from the Conductor team to start Orkes. The company was incorporated in 2021 and introduced its cloud service in 2022. Its proposition preserved the open-source core while offering the managed operations, security and governance that large organizations tend to prefer buying over babysitting.
George describes the boundary in plain terms. Capabilities that directly help developers build should keep flowing into open source. The operational burden around running critical systems - high availability, security, governance, compliance and demanding performance - belongs in the enterprise product. It is a useful founder rule because it is legible. Give the community the instrument. Charge for tuning the hall and keeping the lights on.
The agent enters, improvising
Orkes raised a $20 million Series A in 2024. In April 2026 it announced a $60 million Series B aimed at the growing problem of reliable AI and agentic workloads. The change in language, from microservices to agents, is important but not total. An agent is another distributed application, except now part of the control flow may be probabilistic. It may choose a tool, reconsider, ask another agent, wait for a person and resume much later. The behavior is fresh. The need to preserve state is ancient by software standards.
George sees coding agents as an accelerant. If software becomes easier to create, more of it will be created. The runtime underneath becomes more consequential, not less. A quick demo can survive with a request-and-response loop. A process touching money, customer data or a business operation needs to show its work, retry predictably and fail in a way that can be understood.
This philosophy also explains Agentspan, the open-source project Orkes introduced for building and observing agentic applications. The names change as the industry changes. The underlying taste remains stable: keep developers in their preferred languages, make execution inspectable, and refuse to confuse autonomy with amnesia.
The founder without a happy path
George is notably unsentimental about company building. He has said that startup coverage tends to show the happy path, while the actual work contains non-technical hurdles, misalignment, distractions and days that bear little resemblance to the pitch deck. “Startups are not for everyone,” he said in one interview. “It's a very, very long journey that you're signing up for.”
He sounds less like a motivational poster than an engineer reviewing a failure mode. Founders can focus on the wrong problem. Teams can pull apart. A company can refuse the pivot it needs. The defense is not permanent optimism. It is attention: the right people for the current stage, direct contact with customers, and a long-term vision protected from whatever has become exciting this week.
There is, however, obvious pleasure in the work. George has said there is nothing else he would rather be doing. His ambition is expansive - put Conductor in the hands of developers everywhere - but his method is stubbornly operational. Make the process visible. Keep the state. Expect interruption. Design the recovery.
That may be why his career, from mechanical systems to software systems, feels consistent. An engine is interesting because its parts transfer force without forgetting their relationship. A streaming platform, a marketplace and an AI agent are more abstract machines, but abstraction does not excuse loose bolts.
Reliability after the applause
The AI market enjoys theatrical verbs: agents reason, plan, collaborate and act. George's verbs are quieter: persist, observe, retry and recover. They are what happens after the model has impressed the room and before a customer trusts it with anything important.
Orkes now sits at that threshold. Its origin is a Netflix project built for the complexity of distributed services. Its present is a company arguing that agentic software needs the same durable spine. The proposition is almost comic in its sobriety. The machines may improvise. The infrastructure should keep notes.