Agent cost calculator
Estimate repeated video-query cost before agents touch production traffic.
Video agents can become expensive when every plan, retry, or follow-up question rebuilds visual context. Ayneye moves the repeated work onto durable artifacts: World-State, evidence, timeline, detector report, and cost.json.
Agent cost model
| Layer | What happens | Cost meaning |
|---|---|---|
| One-time processing | Create scene_graph, timeline, evidence, cost | Useful before repeated agent questions |
| Repeated agent queries | Read compact state and evidence | Lower repeated video-understanding work |
| Agent reasoning | Text/model/application compute | Still costs money and should be budgeted |
| Human review | Required for risky actions | Not optional in security/identity/safety workflows |
Implementation pattern
state = ayneye.get_scene_graph(video_id)
evidence = ayneye.get_evidence(video_id)
cost = ayneye.get_cost(video_id)
answer = agent.answer(question, context={"state": state, "evidence": evidence, "cost": cost})
assert answer.cites_evidence
assert not answer.performs_destructive_actionSafety and claim boundary
This is not free infinite querying and not feature equivalence with every GPU video model. It is a way to reduce repeated video-understanding work while preserving evidence and review boundaries.