- Runway Characters, launched March 9, 2026, is a real-time video agent API that generates expressive digital personas from a single image with no fine-tuning required, powered by GWM-1.
- Avatars respond in under two seconds with natural lip-sync and facial gestures, and can execute live UI actions and backend data queries through a split client-server tool architecture.
Full-stack developer and YouTube educator Sonny Sangha, founder of PAPA React and creator of the “Zero to Full Stack Hero” community, published a hands-on build tutorial demonstrating Runway’s Characters API. In the video posted to his YouTube channel, Sangha builds a live AI avatar named Nova from scratch; an avatar that answers spoken questions, reads a company dashboard, closes support tickets, and navigates between screens entirely through voice. Nova responds in under two seconds. Its face moves throughout. It is a working local build, not a pre-recorded demo.
AI Agents Have a Face Now
AI agents have spent two years proving they can book appointments, search databases, and send emails. Almost none of them feel like a conversation, because nearly all of them sit inside a text box or a voice widget. Runway launched Characters on March 9, 2026, describing it as a real-time video agent API that gives developers complete control over an avatar’s voice, personality, knowledge, and actions. The pitch is direct: chatbots made interaction more dynamic, but the medium stayed the same. Characters change the medium.
What GWM-1 Actually Does?
The model underneath Characters is not a video generator in the standard sense. GWM-1 is an autoregressive model built on Gen-4.5 that generates frames in real time, driven by audio inputs. GWM-Avatars version generates realistic human body and face motions in order to produce photorealistic or stylized characters with facial expression animation, eye gaze, lip sync, and gesture animation while talking and listening, throughout long conversations without any loss of quality.
The system runs at 24 frames per second and 720p resolution, with GWM-1 released in three branches: GWM-Worlds, GWM-Robotics, and GWM-Avatars, each targeting a distinct domain. Avatars is the branch that powers Characters.
What most coverage of the launch missed: GWM Avatars maintains natural idle behavior, subtle movement, and responsive expressions even when the character is not speaking, so the conversation feels less like watching a recording and more like a genuine exchange. Idle behavior is what makes a video call feel real. It is also what distinguishes a believable avatar from a pre-baked clip cycling on a loop.
How the Demo Works?
Sangha builds a fictional company dashboard with revenue data, support tickets, and a settings screen. Nova appears as a corner popup. Share your screen, click start, and Nova immediately begins reading the data, narrating dips in revenue, explaining ticket statuses, closing a ticket on request, and navigating to a different screen when asked. The face moves the entire time.
Two separate interaction types carry the capability. Sessions run as live WebRTC connections, with each session linking a user to an avatar for a single interaction capped at five minutes. Client tools handle frontend actions: highlighting elements, scrolling, navigating views. Server tools handle secure backend calls, such as fetching revenue figures from a protected API. Sensitive data stays on the server; the avatar still gets to reason over it. The architecture is similar in logic to how AI tools are now being used to query live analytics data directly, a pattern The Query Post covered in its report on how Claude’s GA4 integration is replacing manual exports for SEO reporting.
Getting started follows three steps. Supply a single reference image to define the character, configure voice, personality, knowledge base, and conversational actions, then deploy through the API. No fine-tuning is required, and the system supports every visual style, from photorealistic humans to animated mascots to stylized brand characters, from a single image.
Already in Production
BBC and Silverside were using Characters before the public launch date. Silverside had already used Runway to build Coca-Cola’s top-performing ad campaigns at a 65% reduction in production cost compared to a traditional shoot. The BBC relationship suggests broadcast media is moving toward interactive character experiences beyond passive video.
For enterprise teams, an avatar built through Characters can create a support ticket, take customer orders against live inventory, or query an enterprise knowledge base for accurate answers in real time. That is a different category of tool from a voice bot reading a script. It functions closer to a product interface that has a face and a voice.
Two Ways to Build: React SDK or Embed Widget
Sangha demonstrates both integration paths in the tutorial. The React SDK published by Runway on GitHub gives developers full lifecycle control, custom tool definitions, and event subscriptions for complex avatar behavior inside existing applications. The embed widget requires far less code and suits teams that want a conversational layer without building a full frontend integration. According to Runway’s Characters documentation, avatars can also be loaded with internal documentation, product FAQs, or how-to guides through knowledge base integration, giving them context-aware answers rather than generic responses.
Where the Competition Stands?
D-ID, Synthesia, Soul Machines, and Google have each developed human avatar technology for communication and training applications. Runway’s distinction is direct API integration with agent tool-use. The avatars can act on live systems; they do not recite from scripts. That capability is what separates Characters from earlier avatar platforms that produced pre-generated video responses with no real-time decision-making.
Positions on runways, people as part of a greater idea that in the future, the world of online communication will be based on live video: experiences that happen right there on the spot, characters that one can communicate with, and worlds that can be created by people. It remains to be seen whether this perception is true in mass terms, based on the adoption of the API in practice.
What to Check Before Building?
The five-minute session cap per WebRTC connection is a real constraint for anyone planning extended educational conversations or long support sessions. Session reconnect logic belongs in the architecture from the start. Credit costs update with model releases, so the Runway pricing page should be verified directly before scoping any production build.
For product teams deciding on their interface layer, the more useful question is whether a text or voice-only experience is genuinely the right fit, or whether it remains the default simply because video agents were not accessible to most teams until now. With Runway’s Characters API publicly available and Sangha’s complete tutorial and codebase ready to clone from the video description, accessibility is no longer the obstacle. The choice between a chat window and a video agent is now a design decision.
