Agent economics
The economic architecture of video agents is memory, not more retries.
A video-agent product fails when every retry sends more frames, every dashboard refresh rebuilds context, and every support question repeats expensive visual work. Durable World-State changes the cost shape.
Where artifact memory wins
Artifact memory wins when teams ask many questions over the same footage, need evidence citations, need cost reports, and need to debug agent behavior. The system processes video into state once, then agents read that state through narrow, auditable tools.
| Workload signal | Meaning | Recommended path |
|---|---|---|
| High repeat rate | Many questions over the same footage | World-State has strong value |
| Low repeat rate | One-off analysis | Value shifts to evidence and auditability |
| Complex semantics | Subtle behavior or cinematic meaning | GPU/human review may be better |
| Regulated decision | Identity, intent, discipline, emergency | Human review required |
Buyer framework
- Measure processed video hours.
- Count repeated questions per video set.
- Inspect evidence coverage and missed events.
- Check whether answers cite evidence.
- Separate low-risk summaries from high-risk decisions.
- Use cost.json to estimate retention and fallback impact.
Product positioning
Ayneye should be positioned as a low-cost video memory layer for AI agents, not as a magical replacement for every vision system. The strongest promise is practical: make video queryable, auditable, and budgetable.