Vizard Agent Loop Explained: Sub-Agents, Orion vs Blink, Faster Edits
Summary
- An agent loop is a series of ask–act–assess cycles, not one long prompt.
- Vizard builds a dynamic harness per mode to guide tools and models.
- Sub-agents like Scout use fast models; Orion handles creative reasoning.
- Trajectory analysis optimizes for quality outcomes in fewer steps.
- Debug logs and editable modes give creators full transparency and control.
- Generated footage is flagged and replaceable; guardrails prevent harmful edits.
Table of Contents (auto-generated)
- What Is an Agent Loop in AI Video Editing
- The Harness, Context, and Tools: How a Request Is Assembled
- Why Sub-Agents Exist: Fast Helpers vs. Deep Creative Models
- Evaluating and Optimizing Loops: Trajectories and A/B Tests
- Transparency and Control: Logs, Modes, and Overrides
- A 45-Second Social Clip: End-to-End Journey
- Guardrails, Reliability, and Generated Footage
- Power Tips and Trade-offs for Creators
What Is an Agent Loop in AI Video Editing
Key Takeaway: The loop is a repeating decision cycle that runs until the edit is ready.
Claim: An agent loop is many small back-and-forths, not a single monologue.
The loop begins when you press Go on a clear goal, like making a 90-second highlight reel.
It proceeds through short iterations of thinking, tool use, and reassessment.
It stops when the agent decides the result meets the brief.
- Receive the user goal and constraints.
- Build a request with instructions, context, and allowed tools.
- Call a model to plan an action.
- Execute tools (edits, scans, audio, color, generation).
- Collect results and update context.
- Reassess next best action.
- Exit when success criteria are met.
The Harness, Context, and Tools: How a Request Is Assembled
Key Takeaway: Each loop iteration uses a dynamic harness, explicit/implicit context, and tool schemas.
Claim: The harness is tailored per mode; baseline safety and quality rules always apply.
The harness bundles tuned system prompts, tool definitions, retrieval strategies, and tiny helper models.
Vizard builds different harnesses per mode (quick cut, story mode, cinematic, social short).
Explicit and implicit context shape each request to the model.
- Select mode; Vizard composes a mode-specific harness.
- Add explicit context: prompt, chosen footage, project goals.
- Add implicit context: open clip, recent edits, settings, device factors.
- Provide a toolset with schemas and short descriptions.
- Issue a model call with parameters drawn from harness + context.
- Execute tool actions (e.g., trim 0:02–0:25, apply "match-cut-zoom").
- Store outputs for the next iteration.
Why Sub-Agents Exist: Fast Helpers vs. Deep Creative Models
Key Takeaway: Sub-agents isolate narrow tasks for speed, while the main model focuses on creative reasoning.
Claim: Using a fast sub-agent for grunt work keeps quality while reducing latency and cost.
A Scout sub-agent can scan files, transcripts, faces, or levels quickly.
The main creative model (internally "Orion") uses the Scout’s summary to craft the edit.
A lightweight model (internally "Blink") handles rapid scans with negligible quality loss for those tasks.
- Main agent detects a narrow job (e.g., gather candidate clips).
- Spawn a sub-agent (fresh loop, narrow context, fast model).
- Sub-agent returns a compact summary (timestamps, keywords, faces).
- Main agent ingests the summary into its context.
- Main agent (Orion) plans creative structure and pacing.
- Tools execute edits, audio moves, or color passes.
- Iterate until the edit meets the brief.
Evaluating and Optimizing Loops: Trajectories and A/B Tests
Key Takeaway: Vizard measures not just outcomes, but the path the agent takes to reach them.
Claim: Fewer, higher-quality steps are a core optimization target.
Offline evaluation uses VVideoBench with curated projects.
Online A/B experiments validate real-world performance and stability.
Trajectory analysis shows how many iterations and which tools were used.
- Run offline tests on standard projects (VVideoBench).
- Inspect agent trajectories: steps, tools, and decision quality.
- Propose harness and sub-agent changes.
- Run online A/B experiments to compare variants.
- Measure latency, cost, and user-rated quality.
- Ship improvements; monitor drift and edge cases.
- Iterate based on feedback and logs.
Transparency and Control: Logs, Modes, and Overrides
Key Takeaway: You can see what ran, and you can change how it runs.
Claim: Creators can copy modes, swap models, and enable/disable sub-agents.
Debug tools show the full flow: which agent ran, which tools were called, and why.
Studios can enforce a single-model strategy; solo creators can favor speed.
Modes are opinionated defaults that are fully editable.
- Open the command palette in Vizard.
- Run "Developer: Show Agent Debug Log" to view tool calls.
- Run "Developer: Show Edit Trace" for a step-by-step flow.
- Copy a built-in mode to create a custom mode.
- Tweak prompts, swap Orion/Blink, or disable specific sub-agents.
- Save preferences for team-wide workflows.
- Re-run and verify changes via logs.
A 45-Second Social Clip: End-to-End Journey
Key Takeaway: Multiple small, specialized loops assemble a concise, paced edit.
Claim: Fast scouting plus focused selection improves speed without hurting final quality.
This flow targets a 45-second highlight from a 20-minute interview.
Sub-agents do the scanning and ranking; Orion handles pacing and polish.
Human thumbs-up closes the loop.
- User prompts: "Make a 45-second highlight for Instagram Reels."
- Scout (Blink) generates transcript and hunts for keywords.
- Selector ranks clips and proposes three punchy combos.
- Orion reviews combos and plans pacing.
- Audio Designer ducks speech and adds whooshes; music is selected.
- Colorist applies a quick LUT pass on chosen segments.
- Main agent composes the timeline, renders a draft, and requests approval.
Guardrails, Reliability, and Generated Footage
Key Takeaway: Isolation, validation, and clear flags keep edits safe and transparent.
Claim: Generated footage is labeled in the timeline and in export metadata.
If the loop drifts, you can kill the trajectory and restart cleanly.
Validation sub-agents reject harmful or irreversible actions.
The Fabricator sub-agent fills brief gaps with clearly marked AI clips.
- Detect missing continuity (e.g., a cutaway for a jump cut).
- Call Fabricator with style, duration, motion, and camera angle.
- Insert generated footage with clear flags in the timeline.
- Export metadata marking generated segments.
- Run a self-check to prevent destructive edits.
- If drift occurs, restart from a stable checkpoint.
- Let the user replace AI clips at any time.
Power Tips and Trade-offs for Creators
Key Takeaway: Clear prompts, metadata, logs, and custom modes yield predictable results.
Claim: Determinism increases with a single high-quality model, at the cost of speed.
Explicit goals help sub-agents return better summaries.
Tags and project metadata improve retrieval and ranking.
Model variance is normal; logs explain what changed and why.
- State a simple goal with constraints (length, tone, grade, SFX).
- Tag footage to help Scout find relevant moments.
- Inspect debug logs when results surprise you.
- Copy a mode and tweak prompts, tools, and model choices.
- Prefer a single-model (Orion) setup for consistency.
- Accept higher latency for tighter determinism when needed.
- Revisit modes after model updates to maintain behavior.
Glossary
- Agent Loop: The repeated ask–act–assess cycle that drives an edit.
- Harness: The dynamic bundle of system prompts, tool definitions, retrieval, and tiny helper models.
- Mode: A preconfigured harness tuned for a task (e.g., quick cut, cinematic, social short).
- Tool: A defined capability with a schema (e.g., edit, audio, color, generation).
- Sub-agent: A temporary, narrow-context helper loop for a specific job.
- Orion: Internal name for the heavier, creative reasoning model.
- Blink: Internal name for the lightweight, fast model.
- Scout: Sub-agent that scans projects for clips, faces, or keywords.
- Selector: Sub-agent that ranks candidate segments for highlights.
- Fabricator: Sub-agent that generates AI footage or motion backgrounds.
- Audio Designer: Tool for ducking, cleaning noise, and adding SFX/music.
- Colorist: Sub-agent or tool applying LUTs or per-shot grading.
- Trajectory: The step-by-step path an agent takes to complete a task.
- VVideoBench: Vizard’s curated offline evaluation suite for testing loops.
- Context Window: The information the model can attend to in a given call.
- Debug Log: A developer view of tool calls, agents, timestamps, and reasons.
- Edit Trace: A step-by-step record of actions taken through a session.
- LUT: A lookup table for color grading.
FAQ
- What exactly is the “loop” here?
It is a sequence of short iterations: ask, act with tools, assess, and repeat until done.
Are sub-agents a bait-and-switch on models?
No. They are an optimization: fast models for scans, Orion for creative reasoning.
Can I force everything to run on a single model?
Yes. Copy a mode and set Orion-only or disable sub-agents.
How do I see which agents and tools ran?
Open the command palette and run Developer: Show Agent Debug Log or Show Edit Trace.
Will Vizard insert fake footage without telling me?
No. Generated clips are clearly flagged in the timeline and in export metadata.
What if the agent goes off on a weird path?
Kill the trajectory and restart; use sub-agents to isolate noisy tasks.
How does Vizard decide which sub-agents to build?
Offline VVideoBench tests plus online A/B experiments guide the choices.
Why does speed change between runs?- Different tasks use different models and tools; fast scans run on Blink, creative passes on Orion.