Yuval Avidani
Author
Key Takeaway
Agency Agents is a system that transforms IDE coding assistants into specialized AI agent personas with domain-specific expertise. Created by msitarzewski, it solves the problem of generic AI code generation by injecting markdown-based system prompts that define each agent's identity, mission, rules, and success metrics across Claude Code, Cursor, and Aider.
What is Agency Agents?
Agency Agents is an open-source framework that provides 112 specialized AI agent personas for software development. The project agency-agents solves the problem of generic, hallucination-prone AI coding assistants that we all face when trying to get production-ready code from tools like Claude Code, GitHub Copilot, or Cursor.
Instead of asking a generic LLM to "write code", Agency Agents lets us invoke specialized personas like Backend Architect, Security Engineer, Frontend Developer, or Reality Checker. Each agent is defined by opinionated markdown-based system prompts that narrow the LLM's context to a specific domain, drastically reducing hallucinations and enforcing best practices.
The Problem We All Know
We've all been there: We ask our AI coding assistant to implement a feature, and we get code that kind of works but doesn't follow our architecture patterns, ignores security best practices we've spent years establishing, or makes assumptions that break in production. We end up spending more time reviewing and fixing AI-generated code than if we'd just written it ourselves from scratch.
The core issue is that generic LLMs try to be everything to everyone. They lack deep context about our specific frameworks, domain requirements, and quality standards. They might know 100 programming languages but have no understanding of our particular codebase's conventions or the business logic that makes our application unique.
Existing tools like GitHub Copilot or Claude Code are impressive but fundamentally limited by their generic nature. They're trained on the entire internet of code, which means they're mediocre at everything rather than excellent at the specific things we need.
How Agency Agents Works
Agency Agents takes a completely different approach: Instead of one generic assistant, we get an entire agency of specialized experts. The system operates by injecting highly opinionated markdown-based system prompts into our favorite AI coding assistants.
Think of it like this: Instead of hiring one generalist freelancer who kind of knows everything, we're hiring a specialized agency where we can talk directly to the backend expert when we need API work, the security specialist when we need vulnerability assessment, or the frontend expert when we need React components.
The repository includes shell scripts that automatically convert and install these agents across platforms:
- Claude Code and Copilot - injects native .md agent files
- Cursor - generates .mdc rule files that Cursor can read
- Aider - compiles a master CONVENTIONS.md file
Each agent persona includes:
- Identity traits - who this agent is and how they think
- Core mission - what this agent is responsible for
- Domain-specific critical rules - non-negotiable best practices
- Success metrics - how to measure if the agent did a good job
Quick Start
Here's how we get started with Agency Agents:
# Clone the repository
git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents
# For Claude Code or Copilot - inject native agents
./scripts/install-claude.sh
# For Cursor - generate .mdc rule files
./scripts/install-cursor.sh
# For Aider - compile master CONVENTIONS.md
./scripts/install-aider.sh
A Real Example
Let's say we're building a REST API and we want the Backend Architect to design the endpoints. Instead of asking Claude Code "create a REST API", we invoke the Backend Architect agent:
# In our IDE, we activate the Backend Architect persona
# The agent now has context like:
# - REST API design principles
# - Database schema optimization
# - Performance considerations
# - Security requirements
# Our prompt to the agent:
"Design a REST API for a task management system with
users, projects, and tasks. Include proper error handling."
# The Backend Architect responds with:
# - Well-structured endpoint design
# - Proper HTTP status codes
# - Database schema considerations
# - Security measures (auth, rate limiting)
# - Performance optimizations (pagination, caching)
The difference is dramatic. Instead of generic code that we need to refactor, we get architecture-aware solutions that consider performance, security, and maintainability from the start.
Key Features
- 112 Specialized Personas - From Frontend Developers to Security Engineers, DevOps specialists to Reality Checkers who verify if our requirements are actually feasible. Think of it like having an entire software agency on call.
- Multi-Agent Orchestration - We can run complex workflows like a "Nexus Spatial Discovery Exercise" where 8 different agents work in parallel - one designing the API, another handling security review, another optimizing database queries, all coordinated automatically.
- Platform Agnostic - Works across Claude Code, Cursor, and Aider. We're not locked into one platform, and we can use the same agent definitions across our entire team regardless of their preferred IDE.
- Markdown-Based Configuration - Agent definitions are simple markdown files that we can read, edit, and version control. No proprietary formats or complex configurations.
- Production-Ready Output - By narrowing the LLM's context to a specialized domain, hallucinations drop dramatically and code quality matches what we'd expect from a domain expert.
When to Use Agency Agents vs. Alternatives
Agency Agents shines when we're working on complex projects where generic AI suggestions aren't good enough. If we're building a quick prototype or experimenting, standard Claude Code or Copilot might be fine.
Compared to GitHub Copilot or standard Claude Code, Agency Agents requires more upfront setup but delivers dramatically higher quality output. Copilot is great for autocomplete and simple functions, but when we need architecture-level thinking or domain expertise, Agency Agents is in a different league.
Compared to tools like Cursor's Rules for AI, Agency Agents is more opinionated and comprehensive. Cursor lets us define custom rules, but Agency Agents provides 112 pre-built personas with deep domain knowledge. We get both the flexibility to customize and the benefit of expert-crafted agent definitions.
We'd choose Agency Agents when:
- Building production applications where code quality matters
- Working with complex architectures that require domain expertise
- Needing consistent best practices across a team
- Running multi-agent workflows for comprehensive solutions
We'd stick with standard tools when:
- Rapid prototyping where speed matters more than architecture
- Simple scripts or one-off utilities
- Learning a new language where generic suggestions are helpful
My Take - Will I Use This?
In my view, Agency Agents represents what AI coding assistance should have been from day one. The difference between asking a generic LLM to "write code" and consulting with a specialized Backend Architect or Security Engineer is night and day.
I've been frustrated for months with AI assistants that produce code requiring constant refactoring. Agency Agents solves this by giving each agent specific identity, mission, and success metrics. The time saved from not fixing hallucinated code easily pays for the initial setup within the first week.
The multi-agent orchestration capability is particularly powerful. Being able to run 8 specialized agents in parallel - one handling API design, another security review, another performance optimization - feels like having an entire development team working on our codebase.
The limitations are worth noting: We need to understand which agent to invoke for which task, and the initial setup requires learning the markdown prompt structure. There's also a learning curve in understanding how to orchestrate multiple agents effectively.
But for anyone serious about using AI in software development, this is essential infrastructure. The fact that it works across Claude Code, Cursor, and Aider means our entire team can use it regardless of their preferred tools.
Will I use this? Absolutely. The jump in code quality and reduction in hallucinations makes this a no-brainer for production development.
Check out the repository: agency-agents
Frequently Asked Questions
What is Agency Agents?
Agency Agents is a system that injects 112 specialized AI agent personas into IDE coding assistants like Claude Code, Cursor, and Aider, transforming them from generic helpers into domain-specific experts with defined missions, rules, and success metrics.
Who created Agency Agents?
Agency Agents was created by msitarzewski. The project is open-source and available on GitHub for anyone to use and contribute to.
When should we use Agency Agents?
We should use Agency Agents when working on complex production applications where code quality, architecture, and best practices matter more than rapid prototyping speed. It's ideal for teams that need consistent domain expertise baked into their AI coding assistants.
What are the alternatives to Agency Agents?
Alternatives include standard GitHub Copilot (great for autocomplete but generic), Claude Code without agents (powerful but lacks domain focus), Cursor's custom rules (flexible but requires writing rules from scratch), and Aider with custom conventions (works but less comprehensive than Agency Agents' 112 pre-built personas).
What are the limitations of Agency Agents?
The main limitations are the initial setup time, the learning curve in understanding which agent to use for which task, and the need to learn how to orchestrate multiple agents effectively. It's also more complex than needed for simple scripts or rapid prototyping.