It's 2 a.m., and somewhere a coding agent has hit a wall. Maybe it's a legacy codebase that refuses to yield its secrets, or an ambiguous product spec that could mean three different things. The agent does what AI typically does in these situations: halts, or worse, conjures up a plausible-sounding answer that's completely wrong.
Humwork, a tiny San Francisco startup that emerged from Y Combinator's Spring 2026 batch, thinks it has found a better answer. When an agent gets stuck, route it to a live human expert. Fast.
The pitch is deceptively simple: install one command-line tool, and any agent running in Claude, Cursor, or a handful of other platforms can tap into what the company claims is a network of over 3,000 specialists. The company's homepage advertises matching in under 30 seconds, average first replies in under two minutes, and an 83% resolution rate. Then, according to Humwork, the agent picks up where it left off.
Whether that model holds up under real production pressure is another matter entirely.
The Technical Bet
Co-founders Yash Goenka and Rohan Datta—both UC Berkeley graduates who met in the Bay Area startup scene—built Humwork around what's called Model Context Protocol, or MCP. Think of it as a standard plug for AI systems, a way for different tools to talk to each other without custom integrations for every possible pairing.
The advantage is speed. One CLI command, and you're in. Agents running in ChatGPT, Lovable, OpenClaw, or any MCP-compatible platform can suddenly request human backup. The agent hits an edge case, triggers a handoff, gets routed to someone who actually knows the answer, and continues on.
"If it speaks MCP, we support it," the company's site declares—a claim that's both ambitious and, for now at least, difficult to verify at scale.
Goenka, who holds India patent IN 328983 related to graphene supercapacitor manufacturing (a detail that feels almost quaint next to his current venture), and Datta, who previously built an AI voice platform that he says handled over a million minutes of calls, framed the problem in their launch materials as a missing piece of infrastructure. The bottleneck in AI systems, they argue, isn't compute or even context windows. It's knowing when to escalate.
The Expert Economy, Repackaged
Humwork's expert network covers software engineering, design, marketing, product strategy, legal compliance, and finance—basically anywhere an agent might find itself in over its head. Privacy filings from early April reveal that experts go through "voice-based qualification assessments," though the company hasn't disclosed much about what that vetting process actually entails.
The economics are murky. Pricing isn't listed anywhere public, though the terms of service note that "applicable rates are disclosed before a consultation begins." Launch coverage mentioned two consumer-tier plans at $20 and $100, plus an enterprise tier requiring a sales conversation. India Today reported in April, citing Humwork support, that experts earn a flat 70 cents per active minute during chat sessions.
That's... not much. For context, traditional expert networks like GLG or Guidepoint typically charge hundreds of dollars per hour and pay experts a comparable rate. Humwork appears to be betting on volume and speed over depth, which makes sense if you're optimizing for agent handoffs rather than executive consultations.
The privacy documents also note that expert responses may be used for model training or dataset creation, with an opt-out available by emailing the founders. It's a detail that raises questions about whether experts fully understand what they're signing up for—and whether enterprises with strict data governance requirements will be comfortable with that arrangement.
A Growing Field
Humwork isn't alone in this space, though the market is young enough that every player is still figuring out what the market actually is.
RentAHuman.ai launched in February with a focus on physical tasks—delivery verification, field inspections, that sort of thing. HumanOps.io offers similar operator delegation for real-world workflows. Humando.ai markets what looks like a nearly identical MCP-based handoff, though they emphasize scenarios where "your AI needs hands."
The differentiation, at least in theory, is that Humwork targets knowledge work rather than physical execution. An agent can't parse a legacy codebase? Route it to an engineer. Need strategic input on a product roadmap? Find a PM who's seen this before.
Guidepoint—a far larger, more established player—launched its own MCP integration in May, embedding a library of over 100,000 expert interview transcripts into AI workflows. That's not live consultation, but it suggests that even the incumbents see agent interoperability as worth building for.
The Numbers Game
Humwork claims an 83% resolution rate and sub-two-minute first replies, based on metrics displayed on their homepage as of mid-June. Press coverage and a job listing from what appears to have been a pre-launch beta referenced roughly 3,000 questions resolved, though those figures carried no disclosed methodology and may not reflect current performance.
The expert pool has grown from "1,000+" in early coverage to "3,000+" more recently, according to the homepage. That's rapid scaling, which either speaks to strong demand or aggressive recruiting. Possibly both.
But here's the harder question: can sub-30-second matching survive contact with messier, more ambiguous real-world escalations? The beta environment is one thing. Production load, where enterprises expect SLAs and every failed handoff represents actual business risk, is quite another.
Security in the Spotlight
Humwork also launched into a moment when MCP security is under active scrutiny. Industry coverage in June warned about misconfigured servers and prompt-injection vulnerabilities. Recent research flagged runtime faults in MCP deployments that could expose sensitive data or allow unauthorized access.
Humwork's privacy policy emphasizes PII redaction—stripping out personal information before passing context to experts—but the company hasn't published a security whitepaper or third-party audit. For a product that sits in the middle of potentially sensitive agent workflows, that's a gap some enterprise buyers will notice.
What We Know (and Don't)
Humwork is still a two-person operation. The company was founded in 2025 and went through Y Combinator's Spring 2026 batch under partner Tyler Bosmeny. The legal entity is registered as Orange AI Inc. LinkedIn data points to a pre-seed round of $500,000 in October 2025, while Dealroom lists a $125,000 entry in March 2026 that may simply reflect standard YC batch terms. Neither figure is confirmed.
The company posted a senior full-stack TypeScript role in April, a sign that the team is trying to grow—though hiring in this market, particularly for early-stage infrastructure plays, isn't exactly straightforward.
The Larger Bet
Strip away the specific product details, and what Humwork is really betting on is this: as AI agents move from demos to production, the handoff to human judgment won't be a fallback feature. It'll be core infrastructure.
That thesis is hard to argue with. Agents that fail silently or loop indefinitely are worse than useless—they're actively destructive. An agent that knows when it's stuck and can escalate gracefully? That's genuinely valuable.
What remains unclear is whether the economics work. Expert networks are notoriously difficult businesses to scale. Matching speed and quality tend to be inversely correlated. And building a sustainable marketplace where both sides—agents and experts—keep coming back requires balancing incentives that often pull in opposite directions.
For now, Humwork offers engineering teams a practical hedge against the inevitable moments when an LLM can't close the loop. The MCP integration means minimal friction. The promise of expert triage in under 30 seconds addresses a real pain point.
Whether that value justifies the cost—and what that cost actually is at scale—will depend on something the founders can't engineer: whether enterprises trust a two-person startup to sit in the critical path of their AI infrastructure.
That's the bet. We'll see if it pays off.
