Skip to content

Cyborg schema

What's on this page

  • Required and optional fields for persona YAML
  • Security and deployment objects
  • Conventions for Family Office and CRITICAL roles

What this enables

  • Consistent agent authoring and validation scripts

Top-level fields

Field Required Description
job_id yes Stable code (e.g. EXEC0001)
display_name yes Short agent name
human_title yes Full human job title
category yes e.g. AI
sub_category yes EXECUTIVE, ENGINEERING, …
level_band recommended Lx / Mx / Exec
reports_to recommended Manager job_id
direct_reports optional List of job_ids
short_description yes One paragraph
system_prompt yes Full operator prompt
personalities recommended Blend tags
example_phrases recommended Dialogue samples
primary_functionality yes List of focus areas
deterministic_capabilities optional Enum-like strings
llm_capabilities yes e.g. TEXT_GENERATION, REASONING
tags yes Search/filter tags
tools_mcp recommended Allowed tools/MCP servers
knowledge_docs recommended Paths under docs/
job_type yes LLM or DETERMINISTIC
deployment yes namespace, SLA, reliability
resources yes cpu/memory/storage/network
security yes level, authn/z, HITL, invokers
priority yes Scheduling priority
timeout_ms yes Per-call timeout
retry yes Backoff config
dependencies optional Integrations
kpi_owned optional Metrics the persona optimizes
escalation_path optional Human chain

Security object

security:
  level: CRITICAL  # NONE|LOW|MEDIUM|HIGH|CRITICAL
  encryption_required: true
  authentication_required: true
  authorization_required: true
  human_approval_required: true
  allowed_invokers: []  # empty = org default policy

Family Office (PERS*) should set restricted allowed_invokers and CRITICAL level.