Sydney Huang Unveils “Human API,” a Platform That Lets AI Agents Hand Off the Hard Parts to People

The newly launched service positions itself as an infrastructure layer for hybrid AI-human workflows, slotting live operators into the points where fully autonomous software still stumbles.

A new startup led by Sydney Huang (@0xSydney), CEO of Eclipse, is betting that the next phase of artificial intelligence will be defined less by another leap in raw model capability than by how cleanly AI systems can pass work back to people when automation runs out of road.

Eclipse’s new platform, Human API, came out of stealth last month after a period of closed piloting. Its pitch is straightforward. Large language models and AI agents have made measurable progress at reasoning, planning, and generating content, but they still trip over the parts of real workflows that hinge on judgment, verification, or navigating systems originally built for human users. Human API offers a programmatic way for developers to route those moments to a live operator on demand and to receive the result back in a structured form their code can consume.

In materials published alongside the launch, the company framed the gap between AI capability and reliable execution as one of the defining bottlenecks in current computing. The argument is that the web was engineered around human assumptions — visual interfaces, authentication checkpoints, CAPTCHAs, ambiguous UI behavior, edge cases that resist scripting — and that software trying to traverse those surfaces autonomously tends to falter on the last stretch of any given task. The company describes its solution as a “hybrid intelligence layer,” with models doing what they do best and humans inserted wherever reality still demands nuance.

How it works

Rather than competing with traditional outsourcing marketplaces, Human API positions itself as orchestration infrastructure for AI applications. Tasks dispatched through the platform are typed, structured, and addressable through APIs, allowing an agent to escalate to a person much the way it would call any other service. Outputs return in standardized formats that downstream automation can act on without manual review.

Developers can configure confidence thresholds, escalation rules, approval workflows, and fully managed pipelines for jobs that should always involve a human. The company says the aim is not to cap autonomy but to make it more dependable under production conditions, where completion rates tend to matter more than headline model benchmarks.

Examples the company points to include browser-based agents passing off CAPTCHA resolution or account recovery, commerce flows escalating payment verification, research assistants requesting human checks on high-stakes outputs, growth tools running platform-native interactions, and customer support systems that transition mid-conversation from AI to a human representative.

Three design pillars

Eclipse describes the product around three principles. The first, which it calls “structured human execution,” reframes outsourcing as machine orchestration: where labor marketplaces are organized for manual coordination, the platform is engineered so that AI applications can call on human workers programmatically. The second pillar emphasizes reliability over pure autonomy, with the company arguing that fully autonomous prototypes often look strong in demos but break down once exposed to production complexity. The third is what it calls real-world interface coverage — a fallback layer aimed at the long tail of workflows involving identity checks, dynamic interfaces, moderation regimes, and platform trust signals that machines have struggled to read consistently.

Taken together, the framing positions humans less as exceptions to automation and more as a programmable resource sitting alongside models in the execution stack.

Early use cases

In its launch materials, the company said early pilot interest has spanned several categories. Teams building browser agents and workflow automation are using the platform to lift reliability on production deployments. Commerce operators are routing verification-heavy steps through human reviewers. Operations teams are blending AI and human input across customer support, onboarding, and moderation pipelines. Other adopters include applications that need higher-confidence outputs from research and verification flows, and data-collection projects where fully automated extraction remains unreliable.

The company has not disclosed pricing, customer counts, or funding details in connection with the launch.

A bet on hybrid work

Beyond the product itself, the launch reflects a broader bet about how AI-native software will be assembled. Huang and the company argue that the first generation of SaaS connected humans to software, and the next will connect AI systems to humans. In that view, application design shifts from asking whether a model can perform a task end-to-end to asking how a system can be composed for dependable outcomes — with humans stepping in as supervisors, validators, and providers of edge-case intelligence.

Whether that abstraction takes hold will depend on developer adoption and on how the economics of human-in-the-loop execution stack up against continuing improvements in fully autonomous agents. For the moment, Human API is positioning itself for the in-between: a world where models can imagine far more than they can reliably accomplish, and where someone, eventually, has to close the remaining gap.

The company described today’s launch as the start of a longer effort, with additional product updates expected as the platform expands.