Build Systems, Not Code
summary
Claim: engineering moves up a layer — an agentic system called Relocation Scout is built with modular skills and sub-agents, structured outputs written into Compendium Wiki memory, and a lint pass that enforces idempotency so retries don’t double-send emails. The concrete rules are: use code for exact answers (commute math, dedupe), use models for judgment, design workflows with stop/retry/escalate paths, decompose giant prompts into reusable components, and gate risky actions behind human approval.
AI-written summary · about
chapters (12)
- 0:00Rediscovering joy in building
- 1:03Engineering skills in agents
- 2:48Agents as system components
- 3:51Designing agent workflows
- 5:14Decomposing prompts into modules
- 7:43Reusable agent capabilities
- 9:31Code versus agent tasks
- 11:19Structured agent outputs
- 13:06State and idempotency
- 15:31Security for agents
- 17:19Maintainability in agent design
- 19:04Agent building as software engineering
related talks
30mBuilding the systems that build the software
16mDon't Build Agents, Build Skills Instead
21mDesigning in the age of the 10x engineer
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
25mBeyond the IDE: Toward Multi-Agent Orchestration
40mHow To Tame Your Agents?
50mCoevolution of Architecture & Code - Closing The Gap
44mBuilding Agentic Workflows in Cursor IDE