Skip to main content

Skill Registry & Multi-Agent Gate Patterns

This guide extends the AEEF orchestration model with a governed skills layer. It defines how orchestrators validate skills before agent execution and how skill usage is tied to stage gates, human approvals, and audit evidence.

Use this page with:

Why Add a Skills Layer to Orchestration​

The base orchestration model controls agent identity and stage order. A skills layer adds control over what reusable execution units agents are allowed to invoke.

Without this layer, a compliant agent can still run an unsafe skill or an unreviewed community workflow.

Core Design Pattern​

Every agent execution request passes through four checks:

  1. Agent Contract Check — Is the agent allowed to perform this class of work?
  2. Skill Registry Check — Is the skill registered, approved, and in policy scope?
  3. Execution Policy Check — Is the skill allowed for this stage, role, and environment?
  4. Gate Binding Check — Are required gates attached before downstream promotion?

Minimum Orchestrator Skill Gate​

skill_gate:
  requested_by_agent: developer-agent
  stage: implementation
  environment: dev
  requested_skill:
    skill_id: aeef.dev.secure-endpoint-template
    version: 1.0.0
  checks:
    agent_contract_allows_skill_category: true
    skill_registered: true
    skill_status: provisional
    role_allowed: true
    environment_allowed: true
    required_gates_attached:
      - code-review
      - testing
      - security-scan
  decision: allow_with_conditions
  human_approval_required: false

Stage-by-Stage Skill Gate Mapping​

StageTypical SkillsRequired ControlsHuman Escalation Trigger
Requirementsstory hardening, acceptance criteria draftingrole scope, prompt/output template checksrisk tier ambiguity
Designarchitecture review, handoff generationarchitect-only skills, pattern allowlistarchitecture-impacting changes
Implementationscaffolding, test generation, refactor assistancecode review + test gates attachedcommand execution outside allowlist
Testingtest matrix, regression analysisevidence logging, non-destructive executionfailing tests with suggested risky fixes
Security/Compliancesecure review, dependency checks, evidence assemblysensitive skill classification + role gatingcritical findings / restricted data
Deploymentmanifest checks, rollout validationproduction-deny by default for skill executionany prod command or infra mutation
Operationstriage summarization, incident templatesread-only runtime tools unless incident moderollback or config mutation

Multi-Agent Handoff Rule for Skills​

When a skill is used, the handoff artifact SHOULD include:

  • skill-id
  • skill-version
  • purpose of use
  • output artifacts produced
  • required gates attached
  • residual risks introduced by skill assumptions

Handoff Extension Example​

handoff_skill_usage:
  skill_id: aeef.test.risk-matrix
  skill_version: 1.1.0
  usage_purpose: generate risk-based integration test matrix
  outputs:
    - tests/risk-matrix.md
  required_gates_bound:
    - testing
    - code-review
  residual_risks:
    - "Missing domain-specific negative tests unless QA agent augments matrix"

Vibe Coding Control Pattern (Agent + Skills)​

AEEF does not ban fast AI-assisted workflows; it requires they are controlled.

Controlled Vibe Coding Pattern​

  1. Developer starts from approved skill catalog.
  2. Orchestrator validates skill scope and required gates.
  3. Developer-agent runs skill with constrained tools.
  4. QA/security skills run as required by gate bindings.
  5. Human review remains mandatory for production-bound code.

Anti-Pattern (Disallowed)​

  • importing community skills directly into production agents
  • letting skills run unrestricted commands
  • skipping review because the skill is "trusted"
  • omitting attribution for externally sourced skills

Enterprise Rollout Sequence​

  1. Add skill registry fields to agent run records.
  2. Enable skill gate validation in dev/staging only.
  3. Require gate binding for implementation and security skills.
  4. Block unregistered skills for production-bound workflows.
  5. Add deprecation enforcement for revoked skills.

Metrics to Track​

  • skill execution count by skill-id
  • skill gate deny rate
  • provisional-to-approved conversion rate
  • incidents involving skill-enabled workflows
  • % of PRs with skill-id traceability metadata
Last updated on by BalrogEg