Build Systems, Not Code - Angie Jones, Agentic AI Foundation
Summary
Angie Jones's "Build Systems, Not Code" posits that designing agentic AI systems is a form of software engineering, shifting the focus from writing code to architecting complex systems. Using "Relocation Scout," a house-hunting agent, as an example, the article demonstrates how traditional engineering skills remain vital. These include systems thinking, workflow design, decomposition of large prompts, separation of concerns, modularity through reusable skills and sub-agents, algorithmic thinking for task allocation, defining contracts for structured outputs, designing for idempotency, and robust threat modeling. The author stresses baking maintainability into the system, ensuring clarity and ease of modification for both human and agentic updates.
Key takeaway
For AI Architects and Software Engineers building agentic systems, prioritize established engineering disciplines over solely prompt engineering. You should design for maintainability by decomposing complex tasks, defining clear contracts for agent outputs, and implementing idempotency to manage state and retries. This approach ensures your agentic solutions are robust, reliable, and easily adaptable, mitigating risks associated with unpredictable model behavior and fostering long-term system health.
Key insights
Architecting agentic systems demands applying established software engineering principles to new AI primitives.
Principles
- Treat agents as components within a larger system.
- Use code for deterministic tasks, agents for judgment.
- Define structured contracts for reliable agent outputs.
Method
Design agents by integrating systems thinking, workflow design, decomposition, separation of concerns, modularity, algorithmic thinking, contracts, idempotency, threat modeling, and maintainability.
In practice
- Decompose "giant prompts" into distinct, manageable jobs.
- Implement structured schemas for agent memory and outputs.
- Design for idempotency to handle retries and state changes.
Topics
- Agentic Systems
- Software Engineering Principles
- System Design
- Workflow Automation
- Prompt Engineering
- AI System Maintainability
- Threat Modeling
Best for: AI Engineer, Software Engineer, AI Architect
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.