Real-time AI video crossed a meaningful threshold in early May 2026. Runway shipped Characters, a system that turns a single reference image into a streaming, conversational video avatar at 24 frames per second with about 1.75 seconds of end-to-end latency. A few weeks earlier, Pika launched PikaStream, with a similar focus on live agents that can speak, listen, and respond inside an ongoing call.
What a Real-Time AI Video Agent Is
A real-time AI video agent is a generative system that produces video output continuously, frame by frame, in response to live input. The defining constraints are:
- Latency: response in under two seconds from input to output
- Frame rate: at least 24 fps, sustained
- Persistence: the same character identity, voice, and style across an entire session
- Bidirectionality: the system listens and responds, instead of rendering a static prompt once
Runway describes Characters as taking a single reference image and producing fully expressive, conversational video with about 37 ms of model time per frame and roughly 1.75 seconds turnaround from when the user stops speaking. PikaStream and Equos target the same shape of problem: a face you can call, with a voice and a personality, that responds at conversational speed.
What These Agents Are Good For
The natural use cases:
- Customer support and sales: an avatar that handles inbound calls, demos, or onboarding
- Tutoring and learning: a tutor avatar a student can interact with at any time
- Game and product NPCs: characters with persistent identity inside a real-time experience
- Accessibility and translation: live conversational interfaces for users who prefer face-to-face interaction
These workloads share a common shape: live, two-way, ephemeral. The video is not edited, mastered, or distributed. It exists during the conversation and disappears after.
Why Production AI Video Editing Is a Different Problem
Production AI video creation — the kind of video you publish, embed, and reuse — has different requirements:
- Frame-accurate control: trimming, retiming, and stitching at the exact frame
- Multi-clip editing: combining generated and uploaded footage on a timeline
- Multi-layer editing: showing videos on top of other videos
- Overlays and motion graphics: adding visual elements on top of your videos like titles, effects, schemas
- Audio and visual mastering: levels, colour, transitions, B-roll, captions
- Reproducibility: the same project can be re-rendered, exported in multiple aspect ratios, and refined later
- Multi-model orchestration: picking the right generation/search model for each shot
A 1.75-second response latency is a signature feature of a real-time avatar. It is irrelevant for an edited explainer video, where the deliverable is a single MP4 that has been iterated on many times. The two systems make opposite trade-offs: real-time agents prioritise immediacy at a fixed quality bar, while production editors prioritise quality, control, and reusability across many models.
Where Tellers Fits
Tellers is a production AI video editing platform. We do not stream avatars on a phone call. What Tellers does:
- Edit on a real timeline: an AI agent that works with frame-accurate cuts, generated clips, B-roll, and transitions
- Edit real footage: the agent can search through your holidays videos or through thousands of hours of source footage for a TV show
- Multi-model generation: Seedance 2, Runway Gen-4.5, Kling, Veo 3.1, LTX, Hailuo, and more, inside one workflow
- API and MCP access: every capability is reachable from your code or your AI assistant
- Reproducible projects: timelines you can revisit, branch, re-render, and export
Both Categories Will Grow
Real-time agents and production editing are not competitors. Most companies will use both for different jobs. A startup launching a product might use a real-time avatar for inbound onboarding, and a production editor to ship a 60-second launch reel for the website and YouTube — same brand, two pipelines.
The interesting infrastructure question over the next year is how these categories integrate. A real-time conversation could be captured, indexed, and turned into edited highlights inside a production workflow. An edited explainer could serve as the fallback when an avatar’s connection drops. Real-time and production AI video are converging on a shared substrate of generation models, but the user-facing tools will stay specialised.
What is Runway Characters?
A real-time AI video system, announced by Runway in early May 2026, that turns a single reference image into a streaming conversational avatar at 24 fps in HD, with about 1.75 seconds of end-to-end latency from speech to video response.
Is Tellers a real-time AI video agent platform?
No. Tellers is a production AI video editing platform. It is designed for videos that get exported, embedded, and distributed as files — not for live conversational avatars on a phone call.
What is the difference between a real-time AI video agent and AI video editing?
Real-time agents generate video continuously in response to live input, optimised for sub-two-second latency. AI video editing is iterative, frame-accurate, and multi-clip — optimised for the quality and control of a deliverable.
Can I use a real-time avatar's output inside an edited video?
Yes, if the platform lets you record or export the session. You can then upload the footage to Tellers and edit it on a timeline alongside generated clips, B-roll, and other assets.
When should I use real-time vs production AI video?
Use a real-time agent when the experience is the product — support, tutoring, NPCs, live demos. Use a production editor when the deliverable is a video file you need to publish, embed, or reuse across channels.
If you have a story to tell or never-edited footage then you should meet Tellers’ agent! open Tellers.