v4.0 Release

Agents That Know
Their Own Job

Every agent declares what skills it uses, what knowledge it reads, and when. Mandatory enforcement sections ground every response in authoritative sources. No more agents winging it from training data.

133 skills · 28 knowledge packs · 40+ skills per agent · 100% metadata coverage

Agents
13 → 12
Skills
133 → 132
Knowledge Packs
10 → 28
Skills / Agent
3–5 → 40+
🧠
Architecture

Agents That
Declare

v4 restructures every agent as a self-describing skill under the Agent Skills Specification standard. Agents declare their core skills, supporting skills, mandatory invocations, knowledge packs, delegation patterns, and collaboration maps — all in structured metadata that tools can read.

  • ✓Every agent lists its core skills, supporting skills, and RACI in metadata
  • ✓Every skill declares its owner and which agents consume it
  • ✓One PRINCIPLES.md per subsystem, inherited by all its agents
  • ✓Agent Skills Specification compliant — works across tools
v4.0 Architecture
🏗️
Three Tiers

Three Tiers of
Knowledge

Agents consult authoritative sources before producing output. Three levels of knowledge — always-loaded, conditional, on-demand — replace the old "reference as needed" pattern that LLMs regularly ignored.

  • ✓Tier 1 — Always loaded. PRINCIPLES.md + preload packs read before every response
  • ✓Tier 2 — Conditional. Keywords like "M&A" trigger blocking reads automatically
  • ✓Tier 3 — On-demand. 28 packs available when the agent needs domain depth
  • ✓Self-Check — Every response verified before output. Non-compliant = invalid
Enforcement-First
📊
Depth

40+ Skills Per Agent.
Declared.

v4 agents ship with 40+ declared skills, 28 knowledge packs, and 100% metadata coverage. 20 cross-domain specialist skills in legal, HR, design, customer success, architecture, and marketing are now available to every agent. Every skill table includes meaningful trigger descriptions — not generic labels.

  • ✓133 skills — each with trigger descriptions, not generic labels
  • ✓28 knowledge packs — up from 10 in v3. Legal, HR, architecture, marketing, finance, ops
  • ✓40+ skills per agent — up from 3–5 in v3. Every agent has a full declared toolkit
  • ✓20 cross-domain skills — risk analysis, contract review, compliance, AI audit, design briefs
  • ✓Mandatory invocations — specific task types that MUST trigger a skill before output
10x More Depth Per Agent
💻
Concrete

See It
In Action

Here's what a v4 agent looks like. This is the Product Manager's metadata and enforcement sections — every one of the 12 agents follows the same structure.

metadata: core_skills: [prd, feature-spec, user-story, ...] mandatory_skill_invocations: - skill: prd triggers: Any PRD authoring escape: quick scoping → /prd-outline conditional_knowledge_packs: - pack: user-research.md trigger_keywords: user research synthesis
## MANDATORY FIRST ACTIONS 1. If matter involves user research → Read user-research.md 2. For Any PRD authoring → MUST invoke /prd 3. For Feature touching personal data → MUST invoke /privacy-policy-audit ## Self-Check Before Output - [ ] Did I read required packs? - [ ] Did I invoke mandatory skills?
@pm Example

One Link. Full System.

Point your Claude Code, Codex, Cursor, Copilot, Gemini CLI, or Windsurf agent at the guide. It handles the rest.

github.com/yohayetsion/product-org-os/.../agent-guide.md