Skip to content

andrewvaughan/agent-council

type overview
description Multi-perspective AI development workflows using councils of specialized agents.

Agent Council

Ship better code by replacing single-agent guesswork with structured, multi-perspective AI workflows.

When a single AI agent plans, builds, and reviews its own code, blind spots compound. Agent Council fixes this by organizing 13 specialized agent personas into 6 councils — each bringing a distinct perspective (security, architecture, quality, operations, user experience) to every significant decision. The result: fewer missed edge cases, stronger security posture, and production-ready code that's been evaluated from multiple angles before it ever reaches a pull request.

Tip

New to council-driven development? Jump to Why Councils? to understand the approach, or skip straight to Installation to try it.

Skills

Nine executable workflows covering the full development lifecycle:

Skill Phase What it does
plan-feature Plan Product Council evaluates scope and priority; Feature Council produces a technical plan with decision records
build-feature Build Full-stack implementation across database, API, frontend, and tests — with user checkpoints at every layer
build-api Build Backend-focused API and database development; activates Architecture Council for significant API decisions
review-code Review Automated SAST scanning + 4-member Review Council for security, quality, documentation, and domain review
submit-pr Ship Pre-submission quality checks, PR creation with conventional formatting, and CI pipeline monitoring
hotfix Fix Streamlined urgent-fix pipeline — branches, implements, focused review, creates PR, monitors CI
gtm-review Launch Go-to-market readiness review combining performance, accessibility, security, content, and GTM Council vote
security-audit Audit SAST + STRIDE threat modeling + attack tree analysis — standalone or as part of the review pipeline
setup-design-system Design Initialize or extend a design system with accessible components, Tailwind/shadcn integration, and docs

Councils

Councils are groups of agent personas that evaluate proposals through structured voting. Each member votes Approve, Concern, or Block with rationale — surfacing risks and trade-offs that a single perspective would miss.

Council Members When it activates
Product 6 agents Feature planning — evaluates scope, priority, and user value
Feature 4 agents Technical planning — breaks down implementation and identifies risks
Architecture 4 agents API design — reviews schema changes, patterns, and scalability
Review 4 agents Code review — assesses security, quality, testing, and documentation
Deployment 3 agents PR submission — checks release readiness and infrastructure impact
GTM 5 agents Launch readiness — content accuracy, performance, accessibility, and go-to-market readiness

Agent Personas

Thirteen specialized perspectives that staff the councils:

Agent Focus
Principal Engineer Architecture, system design, cross-cutting concerns
Security Engineer OWASP, threat modeling, auth, secrets management
QA Lead Test strategy, coverage, edge cases, regression
Frontend Specialist Components, accessibility, state management, UX
Backend Specialist API design, services, data access, performance
Platform Engineer Infrastructure, deployment, monitoring, reliability
DevX Engineer Documentation, developer experience, onboarding
Product Strategist Product vision, user needs, feature prioritization, go-to-market
Lean Delivery Lead Rapid iteration, feature flags, incremental delivery, shipping speed
Design Lead UI/UX design, design systems, accessibility, usability
Business Operations Lead Cost management, ROI analysis, resource allocation, operational impact
Content Reviewer Marketing content accuracy, messaging alignment, legal compliance
Performance Analyst Frontend bundle analysis, Core Web Vitals, backend API performance

Each agent has complexity tiers (Standard and Advanced) that scale review depth based on the significance of the decision.

Why Councils?

A single AI agent reviewing its own work is like a developer merging their own PRs without review — it works until it doesn't. Council-driven development addresses three fundamental problems:

Perspective diversity. A security engineer spots injection risks that a frontend specialist wouldn't flag. A QA lead identifies untested edge cases that a backend developer considers obvious. By routing decisions through multiple specialized perspectives, councils catch the issues that single-agent workflows systematically miss.

Structured checkpoints. Each skill enforces explicit approval gates — you approve the schema before migration runs, the API contract before frontend builds against it, the review findings before they're committed. The AI proposes; you decide. This keeps the human in control without requiring you to micromanage every line.

Audit trail. Council evaluations, votes, and rationale are recorded in decision documents. When you revisit a feature six months later and wonder "why did we design it this way?", the council record has the answer — complete with trade-offs considered and alternatives rejected.

graph LR
    A[plan-feature] --> B[build-feature]
    A --> C[build-api]
    B --> D[review-code]
    C --> D
    D --> E[submit-pr]
    F[security-audit] -.-> |standalone| D
Loading

Skills enforce strict scope boundaries — each owns a single phase and hands off to the next. build-feature never creates a PR. review-code never pushes code. This prevents scope creep and makes each step independently auditable.

Installation

npx skills add andrewvaughan/agent-council

This installs self-contained skill packages — each with its own bundled agent definitions and council templates — into your project's skills directory.

Post-Install

Note

Claude Code only. The steps below apply specifically to Claude Code. For other AI coding agents (Cursor, Codex CLI, etc.), see the Agent Compatibility guide.

Agent Council skills invoke specialized subagents from the wshobson/agents plugin marketplace. Install the plugins before using the skills:

Step 1 — Add the plugin marketplace (one-time):

/plugin marketplace add wshobson/agents

Step 2 — Install each plugin (or use /plugin to open the interactive manager and install from the Discover tab):

/plugin install code-documentation@claude-code-workflows
/plugin install debugging-toolkit@claude-code-workflows
/plugin install git-pr-workflows@claude-code-workflows
/plugin install javascript-typescript@claude-code-workflows
/plugin install full-stack-orchestration@claude-code-workflows
/plugin install frontend-mobile-development@claude-code-workflows
/plugin install backend-development@claude-code-workflows
/plugin install backend-api-security@claude-code-workflows
/plugin install frontend-mobile-security@claude-code-workflows
/plugin install unit-testing@claude-code-workflows
/plugin install code-review-ai@claude-code-workflows
/plugin install api-testing-observability@claude-code-workflows
/plugin install ui-design@claude-code-workflows
/plugin install security-scanning@claude-code-workflows
/plugin install database-migrations@claude-code-workflows
/plugin install database-design@claude-code-workflows
/plugin install cloud-infrastructure@claude-code-workflows
/plugin install documentation-generation@claude-code-workflows
/plugin install dependency-management@claude-code-workflows
/plugin install deployment-validation@claude-code-workflows
/plugin install observability-monitoring@claude-code-workflows
/plugin install application-performance@claude-code-workflows
/plugin install performance-testing-review@claude-code-workflows
/plugin install seo-technical-optimization@claude-code-workflows
/plugin install seo-content-creation@claude-code-workflows
/plugin install content-marketing@claude-code-workflows
/plugin install accessibility-compliance@claude-code-workflows
/plugin install business-analytics@claude-code-workflows

Skills gracefully degrade when plugins are unavailable — they fall back to manual checklists — but the automated integrations improve quality significantly.

Point your AI agent to AGENTS.md in the installed package for full instructions on available skills, councils, and agents.

Configuring Your Tech Stack

Implementation skills (build-feature, build-api) are technology-agnostic — they adapt to whatever your project uses. Add a ## Tech Stack section to your project's AGENTS.md so skills know what frameworks and tools to target:

## Tech Stack

- Frontend: React with TypeScript
- Backend: Express with TypeScript
- ORM: Prisma with PostgreSQL
- UI: Tailwind CSS + shadcn/ui
- Validation: Zod
- Test runner: Vitest
- E2E: Playwright
- Package manager: npm

Note

If the tech stack isn't defined, implementation skills will stop and ask you what your project uses before proceeding. They'll also suggest updating AGENTS.md when they introduce new technologies, so the stack definition stays current.

Quick Start

1. Plan — Run plan-feature with a feature description. The Product Council (6 members) evaluates scope and priority, then the Feature Council (4 members) creates a technical implementation plan. You approve at each checkpoint.

2. Build — Run build-feature (full-stack) or build-api (backend-only) with the issue number from step 1. The agent implements across all layers with checkpoints for schema approval and API contract review.

3. Review — Run review-code for automated SAST scanning and a 4-member council review covering security, quality, testing, and documentation.

4. Ship — Run submit-pr to push, create a PR with conventional formatting, and monitor CI until green.

Agent Compatibility

Agent Support Notes
Claude Code Full Native skill support, slash commands, Task subagents for parallel council evaluation
Cursor Partial Load AGENTS.md as context, invoke skill workflows manually
Codex CLI Partial Reference AGENTS.md, follow skill steps manually
Other agents Manual Any agent that reads markdown can follow the structured workflows

Skills are plain markdown — no proprietary format, no runtime dependencies. Any AI coding agent that can read files and follow instructions can use them. For detailed per-agent setup instructions, workarounds, and a full feature matrix, see the Agent Compatibility guide.

Customization

Tailor the workflows to your project:

  • Add an agent — Create a .md file in canonical/agents/ following the existing persona format
  • Modify a council — Edit the council template in canonical/councils/ to add or remove members
  • Customize a skill — Edit the workflow in canonical/skills/ and rebuild

For a comprehensive guide including tech stack configuration, agent persona customization with examples, and council composition changes, see the Customization Guide.

Build System

All source content lives in canonical/ — skill workflows, agent definitions, council templates, and shared templates. The build script reads scripts/skill-manifest.json to determine which agents, councils, and templates each skill needs, then assembles self-contained packages in skills/.

# Regenerate skills/ from canonical sources
scripts/build.sh

# Verify generated files match committed files (used by CI)
scripts/build.sh --check

After editing anything in canonical/, run scripts/build.sh to regenerate. The --check flag compares generated output against the committed skills/ directory and reports any drift — CI runs this on every pull request.

Important

Never edit files in skills/ directly — they are overwritten by the build. The .claude/skills/ entries are symlinks to skills/, so editing through them also modifies generated files. Always edit sources in canonical/.

See CONTRIBUTING.md for full details on adding skills, agents, and councils.

Documentation

Document Purpose
AGENTS.md Full reference for AI agents — skills, councils, agents, configuration rules
Skills Reference Detailed per-skill documentation with step summaries and example output
Agent Compatibility Feature matrix and per-agent guidance for Claude Code, Cursor, Codex, and more
Customization Guide How to adapt skills, agents, and councils for your project
CONTRIBUTING.md How to add skills, agents, councils, and submit changes
SECURITY.md Vulnerability reporting and security policy

Model Selection

Skills use different Claude models calibrated to task complexity. Pass the appropriate model to the Task tool's model parameter when invoking subagents:

Role Tier Model
Strategic planning, architecture decisions Advanced claude-opus-4-6
Full implementation, council reviews Standard claude-sonnet-4-6
Focused checks, formatting, quick tasks Efficient claude-haiku-4-5-20251001

For the full guide including per-agent tier assignments, see .claude/README.md.

Contributing

We welcome contributions! See CONTRIBUTING.md to get started.

License

MIT

Acknowledgments

About

Multi-perspective AI development workflows using councils of specialized agents. Plan, build, review, and ship — with 13 agent personas across 6 councils.

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages