AI EngineerJune 25, 202619m

Build Systems, Not Code - Angie Jones, Agentic AI Foundation

TL;DR

Agents are part of a system, not the system itself. Systems thinking means designing the whole environment, including files, tools, humans, and other agents, before writing any code.
Giant prompts are the code smell of agentic systems. Decompose prompts into separate skills, sub-agents, and scripts to avoid drift and improve maintainability.
Use code for determinism, agents for judgment, humans for authority. Tasks with exact answers (like commute calculation) should be plain code; agents handle fuzzy reasoning; humans approve critical actions.
Structured contracts between agent components are essential. Outputs should be in a queryable schema (e.g., scores, reasons) so downstream steps can reliably act on them.
Design for idempotency to handle retries safely. Agents must log completed actions to memory so that retries only finish missing steps, not duplicate work.
Threat model your agent's inputs and actions. Treat all external content as untrusted, and wall off high-risk actions (like emailing sellers) behind human approval.

The Breakdown

Angie Jones argues that the thrill of building returns when you stop using agents to write code and start architecting agentic systems. She walks through designing a house-hunting agent, showing how traditional software engineering skills like decomposition, modularity, and threat modeling apply directly to agent design.