An AI coding agent encounters a bug it cannot solve. It tries once, then twice. The error persists—something about a deprecated API endpoint that the model's training data never quite captured. But instead of spinning in circles or confidently suggesting the wrong fix, the agent does something that still feels slightly surreal: it calls for human backup.
Within half a minute, a software engineer somewhere—verified, paid, online—is reviewing the agent's full context. The code. The error logs. The failed attempts. A brief exchange happens in real time. The agent gets unstuck, the work continues, and nobody had to stop what they were doing to manually escalate a ticket through Slack or email.
This is the pitch behind Humwork, a Y Combinator-backed startup that went live in mid-April with a premise that sounds simple but cuts to one of the thorniest questions in the agent economy: what happens when your autonomous system hits the wall?
The answer, according to founders Yash Goenka and Rohan Datta, is infrastructure—specifically, infrastructure that treats human expertise as a callable service, not an afterthought.
MCP as the Gateway
Humwork operates as what's known as an MCP server. MCP—the Model Context Protocol—is the emerging standard that allows AI agents to call external tools and services, a kind of universal adapter for the agent ecosystem. Any MCP-compatible agent can, in theory, ping Humwork when it needs help. The company says setup takes about a minute: plug in the server, make an API call, or install a plugin.
When an agent requests assistance, Humwork routes the full context—code, documents, errors, sometimes entire project histories—to a matched expert. The handoff is designed to be seamless, and the company says it redacts personally identifiable information before sharing, though the technical details of that claim remain somewhat opaque.
The platform works with tools like Claude Code, Cursor, Lovable, and Replit, as well as broader platforms including ChatGPT, Claude, and Gemini. "If it speaks MCP, we support it," the company states on its website. These integrations appear to be technical compatibility rather than formal partnerships—an important distinction in an ecosystem where strategic alliances often carry weight.
The company has shared some early performance data: an 87% resolution rate and average first reply times under two minutes, as of late May. During a beta phase that preceded the public launch, Humwork says it resolved nearly 3,000 questions. Whether those numbers hold as the platform scales is, of course, an open question.
The Expert Economy, Inverted
What makes the model unusual is that Humwork has built what it calls an "Agent-to-Person" marketplace. The economic relationship is inverted from the typical freelancer platform. Here, the AI decides when it needs a human, not the other way around.
The company claims a pool of more than 1,000 verified experts—engineers, designers, lawyers, marketers—though that figure has grown from earlier reporting and lacks independent verification. These experts are independent contractors, not employees. According to reporting by India Today, they're paid $0.70 per minute during active consultations.
That rate structure raises questions. Is 70 cents a minute enough to attract top-tier expertise at scale? Does the model work if experts are idle most of the time, waiting for pings? And what happens when an agent escalates a problem that requires, say, 40 minutes of deep debugging rather than a quick answer?
There's also the matter of data. Humwork's Terms of Service, dated April 8, grant the company broad rights to use expert consultation responses to train AI models and create datasets for research and development. It's a clause that could matter to experts weighing whether their work might eventually automate their own jobs—or at least contribute to that outcome.
Two Berkeley Grads and a Bet on Escalation
Goenka and Datta, both UC Berkeley graduates, bring complementary backgrounds to the problem. Goenka is a two-time founder and former AI engineer who built his first large language model startup back in 2021, when the field was considerably less crowded than it is now. Datta previously built an AI voice calling platform that the company says automated over a million minutes of phone calls; he also spent time as a data scientist at MicroGrid Labs.
They're part of Y Combinator's Spring 2026 batch. LinkedIn lists a pre-seed round of $500,000 from YC dated October 2025, though the amount is self-reported and hasn't been confirmed through press releases or SEC filings. For now, it's a team of two, actively hiring for founding engineers and go-to-market roles.
The company is targeting enterprise customers with features like priority expert matching, custom expert pools, and volume pricing. No customers have been named publicly, which is not unusual at this stage but does leave some questions about traction unanswered.
Pricing remains somewhat unclear. India Today reported pricing tiers of $20 and $100 in mid-April, though no pricing page was visible on Humwork's website as of late May. The Terms of Service note that credit packages and fee structures are subject to change—a hedge that suggests the business model is still being refined in real time.
The Broader Context
Humwork arrives at a peculiar moment. The agent economy is maturing, but unevenly. Companies are discovering that autonomous systems sound great in demos and break in unpredictable ways in production. An agent that confidently deploys the wrong code can cost more than one that admits uncertainty.
The idea of human-in-the-loop isn't new—it's a concept borrowed from automation and robotics, where safety-critical systems always retain a manual override. What's different here is the framing: human expertise not as a fallback but as a service layer, something that can be called programmatically, priced per minute, and integrated into the agent's workflow.
Whether that framing holds up depends on execution. Can Humwork maintain quality as it scales? Will experts stick around for sporadic, minute-rate gigs? And will enterprises trust a marketplace model for problems that might involve sensitive codebases or proprietary information?
The startup is betting yes. And maybe the agents themselves are learning what human engineers have always known, even if they'd prefer not to admit it: sometimes you just need to ask for help.
