Okibi Lets You Build AI Agents by Describing What You Want in Plain English

The Macro: The Agent Builder Market Is Getting Crowded and Most of It Is Drag-and-Drop Theater

I have a theory about the AI agent builder market. Most of the products in it are solving the wrong problem. They are making it slightly easier for technical people to build agents when the actual opportunity is making it possible for non-technical people to build them.

The drag-and-drop workflow builders are everywhere now. Relevance AI, Flowise, n8n with AI extensions, Rivet. They give you a canvas, you connect nodes, you configure prompts and tools, and eventually you have something that works. These tools are fine. They are also just visual programming environments, and visual programming has a forty-year track record of not actually making things accessible to non-programmers.

The people who succeed with drag-and-drop agent builders are the same people who would have succeeded writing the code directly. The abstraction helps them move faster, but it does not expand the audience. Your head of customer support is not going to learn what a conditional branching node is. Your operations manager is not going to debug a workflow graph.

Meanwhile, the agent market itself is expanding rapidly. Companies are automating tasks that used to require human judgment: processing invoices, triaging support tickets, updating CRM records, generating reports from unstructured data. The demand is real. The bottleneck is not “can we build an agent for this?” It is “who is going to build and maintain this agent?” Most companies do not have enough engineers to automate every two-hour-per-day task that knowledge workers do.

This is where the “Lovable for agents” framing starts to make sense. Lovable (and Bolt and v0 before it) proved that you can generate functional web applications from natural language descriptions. The output is not perfect, but it is good enough to ship. The same approach applied to AI agents could genuinely change who is able to automate their own workflows.

The Micro: Second-Time YC Founders Who Already Built an AI-First Product

Okibi lets you create AI agents by describing what you want them to do in natural language. No flowcharts, no node configurations, no visual programming. You say “I need an agent that monitors our Zendesk queue and drafts responses for common billing questions” and Okibi builds it. The agent can then interact with your internal software systems to actually do the work.

Mahyad Ghassemi and Saurav Mitra are the founders. This is their second time through Y Combinator. Their first company was SigmaOS, an AI-native web browser that reached over 100,000 users and processed 75,000 LLM requests daily. That is a meaningful track record. They built a product that real people used at scale, and they learned what it takes to make AI interactions feel natural and reliable.

Going from a browser to an agent builder is a bigger leap than it sounds, but the underlying skill is the same: designing interfaces between humans and AI systems. SigmaOS taught them how people actually want to interact with AI tools. Not through configuration panels. Not through prompt engineering. Through normal language, the way you would explain a task to a coworker.

They are a two-person team in San Francisco, part of YC’s Summer 2025 batch. The product is a web application, though the site itself was showing a loading state when I checked, which suggests it is either in heavy development or running a client-rendered architecture that does not play well with web scrapers. Either way, the YC listing is detailed and the team is active.

The competitive landscape here is significant. Zapier has Central, which builds agents on top of Zap workflows. Lindy AI is doing natural language agent creation and has raised serious money. There is an entire wave of “describe your agent” products launching right now. The question is not whether natural language agent building is a good idea. Everyone agrees it is. The question is who executes it best, and execution in this space means reliability. Can the agent actually do what you described, every time, without breaking?

The Verdict

I think the natural language agent builder is the right interface for this market. Drag-and-drop is a dead end for non-technical users. The SigmaOS experience gives Mahyad and Saurav legitimate credibility in building AI-first interfaces, which matters more than most people realize. Understanding how humans interact with AI systems is the hard part. The model capabilities are table stakes.

The risk is differentiation. If Lindy, Zapier Central, and Okibi all let you describe agents in plain English, the moat becomes reliability and integrations, not the interface itself. Can Okibi’s agents reliably interact with whatever internal tools a company uses? Can they handle edge cases without breaking? Can they be debugged and modified without starting from scratch? These are engineering problems, and a two-person team is stretched thin trying to solve all of them simultaneously.

In thirty days, I want to see real agent deployments in real companies doing real work. Not demos, not prototypes. In sixty days, I want to know the failure rate. How often does a natural-language-created agent do something wrong, and what happens when it does? In ninety days, the question is whether Okibi has built enough integrations to be useful across industries or whether it works great for the three tools they have connected and poorly for everything else. The vision is right. The market is right. Now they need to ship fast enough that Lindy and Zapier do not eat their lunch.