The govrn Strategist Team

Internal · 2026-06-07. The "all the skills" capability — a composed team of agents that can actually run a govrn engagement, encoded so the work is repeatable rather than living only in one person's head. These are the skills behind the playbook.

The team

Agent	Lens / role	Owns
ai-governance-strategist	Lead	The whole engagement; three-lens synthesis; the client narrative; orchestrates the analysts.
rationalization-analyst	Lens 1 — IT-Rationalization	Inventory, capability map, duplication, disposition, system-of-record, integration, roadmap, ROI.
security-analyst	Lens 2 — Cybersecurity	NIST CSF 2.0 / CIS v8.1 posture, identity, vuln/CVE, IR/DR, vendor risk, privacy.
ai-risk-analyst	Lens 3 — AI-Governance	AI-BOM, bias, prompt-injection, provenance, HITL, drift, agentic, foundation-model risk, transparency, AI-IR.
assurance-signer	Independent attestation	Reviews evidence and issues attestations — firewalled from delivery/remediation. Read-only.

The definitions live in .claude/agents/*.md — so any contributor running Claude Code in this repo gets the team automatically as subagents. Invoke the lead and let it orchestrate, or call a specialist directly for a single lens.

How it maps to the method

        ai-governance-strategist  (lead — scope, synthesize, narrate)
            │
   ┌────────┼─────────────┬─────────────────┐
   ▼        ▼             ▼                 ▼
 Lens 1   Lens 2        Lens 3          (Stage 5)
 rational  security      ai-risk         assurance-signer
 -ization  -analyst      -analyst        (independent attestation)
 -analyst
   └────────┴─────────────┴── feed findings up to the lead ──┘

Each analyst scores its lens against docs/ASSESSMENT.md; the lead synthesizes into the three-lens posture and the story (leading with the AI-governance gap); the signer issues any third-party attestation independently.

Non-negotiable boundaries (baked into every agent)

• HITL always — monitoring/retrieval autonomous; any action on a client estate is human-gated, client is approver-of-record.
• The firewall — Finding ≠ Attestation; the signer never assesses or remediates what it certifies.
• In check — every control maps to a recognized standard; never assert a standard's currency from memory.
• Honesty — manual is never presented as monitored; immature capabilities are labeled.
• We monitor systems, not people — signals from vendor advisories, public CVE feeds, org logs, and public product/vendor chatter — never individuals.

Running it under PAI

If you run the PAI platform, these personas can also be composed via the Agents skill (bun ~/.claude/skills/Agents/Tools/ComposeAgent.ts) to get unique voices + prosody, and assembled into a delivery team for parallel lens work. The .claude/agents/ definitions here are the portable, Claude-Code-native version that needs no PAI install — so contributors get the capability out of the box.

Best-practice loop (rinse-and-repeat)

Discovery → Implementation → Test → Training → Certifications → Maintenance, each stage run through all three lenses, with the HITL gates from docs/PLAYBOOK.md. The strategist owns the loop; the analysts own their lens within each stage; the signer gates Stage 5. Stage 6 (Maintenance) is where the assessment becomes the living, monitored platform — the AIaaS annuity.