Agentic Patterns: Code Cheap, Test Hard, Hoard Smart
Coding agents like Claude Code make code generation cheap—hoard proven solutions, loop for better code, integrate Git/subagents, prioritize TDD/manual QA, and avoid unreviewed commits to ship higher-quality software faster.
Hoard Reusable Solutions and Embrace Cheap Code for Compound Gains
With coding agents, generating code costs pennies, shifting focus from writing to curating quality—good code still demands review and maintenance. Hoard snippets, patterns, and modules you know work, then recombine them rapidly; agents amplify this by automating assembly, letting you prototype faster without starting from scratch. Use the compound engineering loop: agents generate options, you select and iterate, avoiding technical debt by having agents refactor proactively. This produces superior code by exploring more architectural choices humans overlook, like optimal data flows or edge-case handling.
Anti-pattern to dodge: never push unreviewed agent code to collaborators—always diff, test, and iterate personally to prevent cascading bugs.
Master Agent Loops, Git, and Subagents for Reliable Builds
Coding agents run LLMs in a reasoning loop: chat-templated prompts with system instructions, token caching for efficiency, tool calls (e.g., shell, file ops), and iterative refinement. Pair with Git essentials—prompt agents on core concepts like branches/commits, use them to rewrite history cleanly via interactive diffs. Deploy subagents for scale: Claude Code's Explore subagent scouts codebases; run parallel subagents for multiple tasks; specialist subagents handle niches like testing or docs. Official docs recommend this for complex projects, turning solo devs into orchestrators.
Enforce QA with TDD, Agentic Testing, and Code Walkthroughs
Start every session by running tests first—agents fix failures faster in context. Follow red/green TDD: agents write failing tests (red), implement fixes (green), refactor. For manual QA, task agents with browser automation on web UIs, logging issues via Showboat note-taking. Understand code via linear walkthroughs (e.g., Showboat + Present for step-by-step traces) or interactive explanations like word clouds highlighting key terms. Annotated example: build GIF optimizer with WebAssembly/Gifsicle by prompting for architecture, then follow-ups for perf tweaks.
Appendix prompts boost this: Artifacts for structured outputs, Proofreader for polish, Alt text generation, Podcast highlights extraction—reusable for any agent workflow.
This guide's TOC reveals a full system for agentic engineering, not hype: practical loops yield production code 10x faster when habits stick.