CrewAI guide
Split video-agent work into planner, reviewer, and reporter roles.
CrewAI workflows become safer when one role selects artifacts, another reviews evidence, and a final role writes a grounded summary. No role should convert detections into high-risk decisions.
Role design
| Role | Allowed tools | Blocked actions |
|---|---|---|
| Planner | application wrappers over Ayneye status/artifact reads | No final incident or identity conclusions |
| Evidence reviewer | application wrappers over Ayneye artifacts and evidence-grounded Ask | No unsupported claims |
| Reporter | summarize supported findings | No dispatch, access-control, or disciplinary actions |
| Human reviewer | Approve/reject high-risk interpretations | Keeps accountability outside the agent |
CrewAI implementation sketch
planner = Agent(role="Video workflow planner", goal="Choose relevant artifacts")
reviewer = Agent(role="Evidence reviewer", goal="Validate every claim against evidence_ref")
reporter = Agent(role="Grounded summarizer", goal="Write only supported findings")
# Tools remain read-only:
# fetch_scene_graph(video_id)
# fetch_evidence(video_id, evidence_ref)
# fetch_cost(video_id)
# Task rule:
# If evidence is missing, create a human review item instead of a claim.Evaluation matrix
- Planner selects the correct processed video and artifact bundle.
- Reviewer rejects unsupported or ambiguous claims.
- Reporter cites evidence references in every supported finding.
- Workflow creates review items for risky conclusions.
- No agent role can trigger physical or destructive action.