Agent quickstart

Build the first evidence-grounded video agent in seven controlled steps.

Name: Ayneye
Availability: LimitedAvailability

Start with one video, one question family, one artifact bundle, and read-only tools. The goal is not autonomy; the goal is proving that an agent can answer from evidence and refuse unsupported claims.

Cost Economics Quickstart LangChain CrewAI MCP tools World-State docs

Seven-step path

Step	Action	Pass condition
1	Create beta workspace	Dashboard opens with sandbox limits
2	Create demo workspace or add one safe video	Video record exists
3	Run bounded execution	Artifacts are materialized
4	Fetch scene_graph.json and timeline.json	Objects/events/time spans are readable
5	Fetch evidence_bundle.json	Answerable events have evidence refs
6	Expose read-only agent tools	Tool calls are logged and scoped
7	Ask three repeated questions	Agent cites evidence or returns insufficient evidence

Minimal client shape

import requests

BASE = "https://api.ayneye.com"
API_KEY = "ay_live_or_beta_key"

headers = {"Authorization": f"Bearer {API_KEY}"}
video_id = "vid_demo_001"

scene = requests.get(f"{BASE}/api/product/videos/{video_id}/scene-graph", headers=headers).json()
evidence = requests.get(f"{BASE}/api/product/videos/{video_id}/evidence-view", headers=headers).json()
answer = requests.post(
    f"{BASE}/api/product/videos/{video_id}/ask",
    headers=headers,
    json={"question": "What changed near the entrance?", "require_evidence": True},
).json()
assert answer.get("evidence_refs") or answer.get("state") == "INSUFFICIENT_EVIDENCE"

First-agent guardrail prompt

You are a video evidence assistant.
Use only the provided scene_graph, timeline, evidence_bundle, and cost report.
Every factual claim must cite evidence_ref.
If evidence is missing, return INSUFFICIENT_EVIDENCE.
Do not identify people, infer intent, trigger actions, or make access-control decisions.