Constrained Autonomy: An Agentic Architecture Scaling to Hundreds of Thousands of Creators

The presentation begins by describing the complexity of AI‑assisted video production, which involves script generation, asset matching, editing, multilingual adaptation, and content review. Early attempts using single prompts or isolated tools failed to scale because creators had to switch between fragmented systems.

Core Challenges

Orchestration uncertainty : How should agents collaborate without the workflow deviating?

Fault irrecoverability : If an agent crashes, does the whole task fail?

Quality non‑attributability : When the final video is poor, which stage is responsible?

Design Principle – Constrained Autonomy

Agents are given deterministic boundaries within which they can make autonomous decisions. Conflict resolution and self‑healing mechanisms guarantee reliability while preserving flexibility.

Six‑Layer Collaborative Architecture

The system is organized into six layers, each with a clear responsibility:

Access Layer : Supports mobile, web, CLI, API, and overseas users.

Gateway Layer : CDN acceleration and security; separates an AI‑gateway for streaming model outputs from a traditional API gateway.

Business Layer : Core Harness execution framework, domain agents, idempotent skills, event bus, plus cloud plugins and editing capabilities.

Hybrid Technology Layer : Kubernetes for long‑running services, Serverless for burst traffic, and AgentRun Sandbox for AI safety.

Infrastructure Layer : Message queues, Redis, multiple databases, object storage.

Observability Base : End‑to‑end tracing and monitoring across the entire pipeline.

Core Roles in the Business Layer

The central model consists of four core roles plus two auxiliary roles:

Commander : Decomposes a video request into sub‑tasks (pre‑processing, editing, review, etc.) and publishes them to the event bus.

Event Bus : The sole communication hub; agents never call each other directly, ensuring complete decoupling.

Domain Agents : Vertical agents such as ScriptAgent, EditAgent, AssetAgent, ReviewAgent, each handling its own domain autonomously.

Idempotent Skills : Atomic capabilities attached to agents (e.g., clip editing, timeline sequencing, final composition) that are defined once, are deterministic, and support safe retries via model version locking, random‑seed fixing, and parameter‑hash caching.

Arbiter : Resolves resource conflicts using a Redis distributed lock, guaranteeing consistent arbitration.

State Machine : Persists the full lifecycle of a project, drives the next step on state changes, and enables checkpoint recovery and audit.

High‑Concurrency Engineering

To serve tens of thousands of concurrent creators, three optimizations were applied:

RabbitMQ cluster with fair‑dispatch and basicQos=1 plus manual acknowledgments, ensuring each consumer processes one message at a time and preventing slow nodes from being overloaded.

Stateless agents: all context travels in event messages; results are persisted to databases and caches. If an agent instance crashes, a new instance can resume work from the event bus, and idempotent skills make retries side‑effect free.

Distributed infrastructure: Kubernetes for resident services, Serverless for spikes, and a sandbox for AI safety.

Observability Stack

A three‑dimensional observability system was built:

UBA (User Behavior Analytics) : Real‑time reporting of user actions, waiting times, and errors from the moment a video creation is initiated.

OpenTelemetry : End‑to‑end trace across more than ten processing stages, visualized in Grafana dashboards showing success rates, agent latency, and error distribution.

LangSmith + BI Dashboard : Tracks LLM token consumption, latency, and output quality; aggregates business metrics for a data‑flywheel and records full audit logs.

Quality Assurance – Layered Gatekeeping

To avoid over‑filtering creative content, a three‑layer gate is introduced:

L1 – Basic Compliance : Automatic blocking of illegal, pornographic, or violent material.

L2 – Technical Specification : Enforces resolution, bitrate, etc., but allows manual overrides for high‑value creative content.

L3 – Creative Evaluation : Routes unconventional narratives or compositions to rapid human arbitration; machines only flag and forward.

This structure dramatically reduced false‑positive creative kills while keeping baseline safety incidents at zero.

Algorithmic Optimizations for Content Quality

Style‑Memory Model : Learns a creator’s or director’s rhythm and injects style parameters during generation to reduce homogenization.

Narrative‑Arc Evaluation Skill : Quantifies tension curves in scripts, automatically detecting conflict points, climaxes, and setups to ensure a non‑flat story structure.

Emotion‑Driven Visual Matching : Adjusts color grading and composition based on dialogue sentiment.

The metric focus shifted from raw generation speed to adoption rate and completion rate . High‑adoption videos are decomposed into atomic “effective creation patterns” and fed back into agent templates, while critical story points now include a fast human feedback channel (state WAITING_CREATIVE_INPUT) for co‑creative decision making.

Future Outlook

Two long‑term directions are pursued:

Infinite Canvas Director‑Level Control : Inspired by Figma, the linear timeline is replaced by an unlimited canvas where creators place logical nodes; agents execute the detailed steps, delivering industrial‑grade, director‑level output.

Interactive Real‑Time Co‑Creation : With low‑latency streaming inference and WebRTC, creators can issue commands like “replace opening with cherry‑blossom shot” or “add slow‑motion” and see immediate results. This requires support for streaming events, incremental state sync, and dynamic workflow adjustments in the commander.

In summary, the "constrained autonomy" approach—combining event‑driven design, idempotent skills, distributed engineering, and layered quality gates—produces a stable, flexible Agentic platform that currently powers hundreds of thousands of AI‑generated videos and sets the stage for interactive, real‑time co‑creation.